Skip to content

GitLab

Commit 6805df0e authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into gelu

parents 1fdbe3fe e4584d91
Hide whitespace changes
Inline Side-by-side
Showing with 1051 additions and 208 deletions
Jenkinsfile
View file @ 6805df0e
...@@ -7,7 +7,6 @@ def show_node_info() { ...@@ -7,7 +7,6 @@ def show_node_info() {
echo "NODE_NAME = \$NODE_NAME" echo "NODE_NAME = \$NODE_NAME"
lsb_release -sd lsb_release -sd
uname -r uname -r
cat /sys/module/amdgpu/version
ls /opt/ -la ls /opt/ -la
""" """
} }
......
example/12_reduce/reduce_blockwise.cpp
View file @ 6805df0e
...@@ -33,11 +33,11 @@ constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2; ...@@ -33,11 +33,11 @@ constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
constexpr bool PropagateNan = true; constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false; constexpr bool OutputIndex = false;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType; using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation = using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation; typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation = using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation; typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
using DeviceReduceInstance = DeviceReduceMultiBlock<InDataType, using DeviceReduceInstance = DeviceReduceMultiBlock<InDataType,
AccDataType, AccDataType,
...@@ -247,6 +247,13 @@ int main(int argc, char* argv[]) ...@@ -247,6 +247,13 @@ int main(int argc, char* argv[])
DeviceMem out_index_dev(indicesSizeInBytes); DeviceMem out_index_dev(indicesSizeInBytes);
InElementwiseOperation in_elementwise_op;
AccElementwiseOperation acc_elementwise_op;
std::tie(in_elementwise_op, acc_elementwise_op) =
reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
static_cast<int32_t>(reduce_total_length));
if(args.do_verification) if(args.do_verification)
{ {
ReductionHost<InDataType, ReductionHost<InDataType,
...@@ -261,8 +268,13 @@ int main(int argc, char* argv[]) ...@@ -261,8 +268,13 @@ int main(int argc, char* argv[])
OutputIndex> OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims); hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run( hostReduce.Run(alpha,
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data()); in.mData.data(),
beta,
out_ref.mData.data(),
out_indices_ref.mData.data(),
in_elementwise_op,
acc_elementwise_op);
}; };
std::vector<ck::index_t> i_inLengths; std::vector<ck::index_t> i_inLengths;
...@@ -277,20 +289,19 @@ int main(int argc, char* argv[]) ...@@ -277,20 +289,19 @@ int main(int argc, char* argv[])
auto reduce = DeviceReduceInstance{}; auto reduce = DeviceReduceInstance{};
auto argument_ptr = reduce.MakeArgumentPointer( auto argument_ptr = reduce.MakeArgumentPointer(i_inLengths,
i_inLengths, i_inStrides,
i_inStrides, i_outLengths,
i_outLengths, i_outStrides,
i_outStrides, reduceDims,
reduceDims, alpha,
alpha, beta,
beta, in_dev.GetDeviceBuffer(),
in_dev.GetDeviceBuffer(), nullptr,
nullptr, out_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(), out_index_dev.GetDeviceBuffer(),
out_index_dev.GetDeviceBuffer(), in_elementwise_op,
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)}, acc_elementwise_op);
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
if(!reduce.IsSupportedArgument(argument_ptr.get())) if(!reduce.IsSupportedArgument(argument_ptr.get()))
{ {
......
example/12_reduce/reduce_blockwise_two_call.cpp
View file @ 6805df0e
...@@ -31,13 +31,13 @@ constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2; ...@@ -31,13 +31,13 @@ constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
constexpr bool PropagateNan = true; constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false; constexpr bool OutputIndex = false;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType; using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation = using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation; typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation = using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation; typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
using PassThroughOp = tensor_operation::element_wise::UnaryIdentic<AccDataType, AccDataType>; using PassThroughOp = tensor_operation::element_wise::PassThrough;
using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType, using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType,
AccDataType, AccDataType,
...@@ -184,6 +184,13 @@ int main(int argc, char* argv[]) ...@@ -184,6 +184,13 @@ int main(int argc, char* argv[])
if(beta != 0.0f) if(beta != 0.0f)
out_dev.ToDevice(out.mData.data()); out_dev.ToDevice(out.mData.data());
InElementwiseOperation in_elementwise_op;
AccElementwiseOperation acc_elementwise_op;
std::tie(in_elementwise_op, acc_elementwise_op) =
reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
static_cast<int32_t>(reduce_total_length));
if(do_verify) if(do_verify)
{ {
ReductionHost<InOutDataType, ReductionHost<InOutDataType,
...@@ -198,7 +205,13 @@ int main(int argc, char* argv[]) ...@@ -198,7 +205,13 @@ int main(int argc, char* argv[])
OutputIndex> OutputIndex>
hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims); hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(alpha, in_1.mData.data(), beta, out_ref.mData.data(), nullptr); hostReduce.Run(alpha,
in_1.mData.data(),
beta,
out_ref.mData.data(),
nullptr,
in_elementwise_op,
acc_elementwise_op);
}; };
std::vector<ck::index_t> i_inLengths_1; std::vector<ck::index_t> i_inLengths_1;
...@@ -217,20 +230,19 @@ int main(int argc, char* argv[]) ...@@ -217,20 +230,19 @@ int main(int argc, char* argv[])
auto reduce_1 = DeviceReduceInstance_1{}; auto reduce_1 = DeviceReduceInstance_1{};
auto argument_ptr_1 = reduce_1.MakeArgumentPointer( auto argument_ptr_1 = reduce_1.MakeArgumentPointer(i_inLengths_1,
i_inLengths_1, i_inStrides_1,
i_inStrides_1, i_inLengths_2,
i_inLengths_2, i_inStrides_2,
i_inStrides_2, reduceDims_1,
reduceDims_1, 1.0f,
1.0f, 0.0f,
0.0f, in_1_dev.GetDeviceBuffer(),
in_1_dev.GetDeviceBuffer(), nullptr,
nullptr, in_2_dev.GetDeviceBuffer(),
in_2_dev.GetDeviceBuffer(), nullptr,
nullptr, in_elementwise_op,
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)}, PassThroughOp{});
PassThroughOp{});
if(!reduce_1.IsSupportedArgument(argument_ptr_1.get())) if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
{ {
...@@ -243,20 +255,19 @@ int main(int argc, char* argv[]) ...@@ -243,20 +255,19 @@ int main(int argc, char* argv[])
auto reduce_2 = DeviceReduceInstance_2{}; auto reduce_2 = DeviceReduceInstance_2{};
auto argument_ptr_2 = reduce_2.MakeArgumentPointer( auto argument_ptr_2 = reduce_2.MakeArgumentPointer(i_inLengths_2,
i_inLengths_2, i_inStrides_2,
i_inStrides_2, i_outLengths,
i_outLengths, i_outStrides,
i_outStrides, reduceDims_2,
reduceDims_2, alpha,
alpha, beta,
beta, in_2_dev.GetDeviceBuffer(),
in_2_dev.GetDeviceBuffer(), nullptr,
nullptr, out_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(), nullptr,
nullptr, PassThroughOp{},
PassThroughOp{}, acc_elementwise_op);
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
if(!reduce_2.IsSupportedArgument(argument_ptr_2.get())) if(!reduce_2.IsSupportedArgument(argument_ptr_2.get()))
{ {
......
example/13_pool2d_fwd/pool2d_fwd_common.hpp
View file @ 6805df0e
...@@ -31,16 +31,15 @@ static void pool_host_verify(const Tensor<InDataType>& in, ...@@ -31,16 +31,15 @@ static void pool_host_verify(const Tensor<InDataType>& in,
const std::array<ck::index_t, 2>& in_left_pads, const std::array<ck::index_t, 2>& in_left_pads,
const std::array<ck::index_t, 2>& /*in_right_pads*/) const std::array<ck::index_t, 2>& /*in_right_pads*/)
{ {
const int32_t divider = window_spatial_lengths[0] * window_spatial_lengths[1]; const int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
using ReduceOperation = typename ck::reduce_binary_operator<AccDataType, ReduceOpId>::opType; using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation = typename ck::
reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation = typename ck::
reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
const InElementwiseOperation in_elementwise_op(divider); auto elementwise_ops =
const AccElementwiseOperation acc_elementwise_op(divider); ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
auto in_elementwise_op = std::get<0>(elementwise_ops);
auto acc_elementwise_op = std::get<1>(elementwise_ops);
if constexpr(!OutputIndex) if constexpr(!OutputIndex)
{ {
...@@ -48,7 +47,7 @@ static void pool_host_verify(const Tensor<InDataType>& in, ...@@ -48,7 +47,7 @@ static void pool_host_verify(const Tensor<InDataType>& in,
ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>; ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) { auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOperation::GetIdentityValue(); auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y) for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
{ {
...@@ -86,7 +85,7 @@ static void pool_host_verify(const Tensor<InDataType>& in, ...@@ -86,7 +85,7 @@ static void pool_host_verify(const Tensor<InDataType>& in,
AccDataType, AccDataType,
IndexDataType>; IndexDataType>;
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) { auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOperation::GetIdentityValue(); auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
IndexDataType accuIndex = 0; IndexDataType accuIndex = 0;
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y) for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
......
example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp
View file @ 6805df0e
...@@ -41,9 +41,8 @@ using CLayout = ck::tensor_layout::gemm::RowMajor; ...@@ -41,9 +41,8 @@ using CLayout = ck::tensor_layout::gemm::RowMajor;
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;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using DsReduceOp = ck::Tuple<ck::reduce::Max<ReduceAccDataType>>; using DsReduceOp = ck::Tuple<ck::reduce::Max>;
using DsElementOp = ck::Tuple< using DsElementOp = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>>;
using DGlobalMemOp = using DGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>; ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
...@@ -236,10 +235,14 @@ int main(int argc, char* argv[]) ...@@ -236,10 +235,14 @@ int main(int argc, char* argv[])
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
ReduceAccDataType d_acc = d_reduce_op.GetIdentityValue(); ReduceAccDataType d_acc = d_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
d_reduce_op(d_acc, c_m_n_host_result(m, n)); {
ReduceAccDataType curr_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
d_reduce_op(d_acc, curr_val);
};
d_m_host_result(m) = d_acc; d_m_host_result(m) = d_acc;
} }
......
example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_fp16.cpp
View file @ 6805df0e
...@@ -41,18 +41,15 @@ using CLayout = ck::tensor_layout::gemm::RowMajor; ...@@ -41,18 +41,15 @@ using CLayout = ck::tensor_layout::gemm::RowMajor;
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;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<ReduceAccDataType>; using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add<ReduceAccDataType>; using D1ReduceOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>; using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using UnaryIdenticElementOp = using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>; using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnaryDivElementOp = using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>; using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using UnarySquareElementOp = using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DGlobalMemOp = using DGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd, ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
...@@ -67,7 +64,7 @@ using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_ ...@@ -67,7 +64,7 @@ using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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| SrcDstScalarPerVector| //######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| 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| _MPerBlock| //######| | | | | | | | | | | Operation| 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| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; < Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
...@@ -204,8 +201,8 @@ int main(int argc, char* argv[]) ...@@ -204,8 +201,8 @@ int main(int argc, char* argv[])
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()), auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer())); static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
auto dxs_in_element_op = DxsInElementOp{}; auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOp{M, M}; auto dxs_out_element_op = DxsOutElementOps{N, N};
// do GEMM // do GEMM
auto gemm = DeviceGemmReduceInstance{}; auto gemm = DeviceGemmReduceInstance{};
...@@ -261,14 +258,14 @@ int main(int argc, char* argv[]) ...@@ -261,14 +258,14 @@ int main(int argc, char* argv[])
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
float d0_acc = d0_reduce_op.GetIdentityValue(); auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
float d1_acc = d1_reduce_op.GetIdentityValue(); auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n)); auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
float d0_val = 0; ReduceAccDataType d0_val;
float d1_val = 0; ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val); dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val); dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
......
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View file @ 6805df0e
...@@ -39,16 +39,14 @@ using CLayout = ck::tensor_layout::gemm::RowMajor; ...@@ -39,16 +39,14 @@ using CLayout = ck::tensor_layout::gemm::RowMajor;
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;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<ReduceAccDataType>; using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add<ReduceAccDataType>; using D1ReduceOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>; using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using UnaryIdenticElementOp = using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>; using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using UnarySquareElementOp = using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>; using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using DGlobalMemOp = using DGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd, ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
...@@ -63,7 +61,7 @@ using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatc ...@@ -63,7 +61,7 @@ using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatc
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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| SrcDstScalarPerVector| //######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| 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| _MPerBlock| //######| | | | | | | | | | | Operation| 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| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; < Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on // clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host:: using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
...@@ -206,8 +204,8 @@ int main(int argc, char* argv[]) ...@@ -206,8 +204,8 @@ int main(int argc, char* argv[])
a_element_op, a_element_op,
b_element_op, b_element_op,
c_element_op, c_element_op,
DxsInElementOp{}, DxsInElementOps{},
DxsOutElementOp{}, DxsOutElementOps{},
BatchCount); BatchCount);
if(!batched_gemm.IsSupportedArgument(argument)) if(!batched_gemm.IsSupportedArgument(argument))
...@@ -259,14 +257,15 @@ int main(int argc, char* argv[]) ...@@ -259,14 +257,15 @@ int main(int argc, char* argv[])
{ {
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
float d0_acc = d0_reduce_op.GetIdentityValue(); auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
float d1_acc = d1_reduce_op.GetIdentityValue(); auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
float c_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n)); auto c_val =
float d0_val = 0; ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
float d1_val = 0; ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
UnaryIdenticElementOp{}(d0_val, c_val); UnaryIdenticElementOp{}(d0_val, c_val);
UnarySquareElementOp{}(d1_val, c_val); UnarySquareElementOp{}(d1_val, c_val);
......
example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View file @ 6805df0e
...@@ -42,8 +42,7 @@ using ABDataType = F16; ...@@ -42,8 +42,7 @@ using ABDataType = F16;
using CDataType = F16; using CDataType = F16;
using EltwiseComputeDataType = F32; using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise:: using Add = ck::tensor_operation::element_wise::Add;
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using DeviceElementwiseAddInstance = using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType, ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
......
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
View file @ 6805df0e
...@@ -17,8 +17,7 @@ using ABDataType = F16; ...@@ -17,8 +17,7 @@ using ABDataType = F16;
using CDataType = F16; using CDataType = F16;
using EltwiseComputeDataType = F32; using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise:: using Add = ck::tensor_operation::element_wise::Add;
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using DeviceElementwiseAddInstance = using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType, ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
......
example/19_binary_elementwise/elementwise_add_1d.cpp
View file @ 6805df0e
...@@ -42,8 +42,7 @@ using ABDataType = F16; ...@@ -42,8 +42,7 @@ using ABDataType = F16;
using CDataType = F16; using CDataType = F16;
using EltwiseComputeDataType = F32; using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise:: using Add = ck::tensor_operation::element_wise::Add;
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using DeviceElementwiseAddInstance = using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType, ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
......
example/19_binary_elementwise/elementwise_add_4d.cpp
View file @ 6805df0e
...@@ -42,8 +42,7 @@ using ABDataType = F16; ...@@ -42,8 +42,7 @@ using ABDataType = F16;
using CDataType = F16; using CDataType = F16;
using EltwiseComputeDataType = F32; using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise:: using Add = ck::tensor_operation::element_wise::Add;
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using DeviceElementwiseAddInstance = using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType, ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
......
example/20_convnd_bwd_weight_xdl/CMakeLists.txt
View file @ 6805df0e
add_example_executable(example_convnd_bwd_weight_xdl convnd_bwd_weight_xdl.cpp) add_example_executable(example_convnd_bwd_weight_xdl convnd_bwd_weight_xdl.cpp)
target_link_libraries(example_convnd_bwd_weight_xdl PRIVATE conv_util) add_example_executable(example_convnd_bwd_weight_xdl_bf16_splitk convnd_bwd_weight_xdl_bf16_splitk.cpp)
\ No newline at end of file target_link_libraries(example_convnd_bwd_weight_xdl PRIVATE conv_util)
target_link_libraries(example_convnd_bwd_weight_xdl_bf16_splitk PRIVATE conv_util)
\ No newline at end of file
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl.cpp
View file @ 6805df0e
...@@ -297,52 +297,15 @@ int main(int argc, char* argv[]) ...@@ -297,52 +297,15 @@ int main(int argc, char* argv[])
split_k); split_k);
// alloc work space // alloc work space
size_t bwd_weight_workspace_size = conv->GetWorkSpaceSize(argument.get()); float ave_time = 0.f;
float ave_time = 0.f; if(!conv->IsSupportedArgument(argument.get()))
if(std::is_same<InDataType, ck::bhalf_t>::value && split_k > 1)
{ {
DeviceMem wei_work_space_device_buf(bwd_weight_workspace_size); std::cout << "wrong! device_conv with the specified compilation parameters does "
wei_work_space_device_buf.SetZero(); "not support this Conv problem"
argument = conv->MakeArgumentPointer( << std::endl;
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()), return 1;
static_cast<AccDataType*>(wei_work_space_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{},
split_k);
if(!conv->IsSupportedArgument(argument.get()))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
}
else
{
if(!conv->IsSupportedArgument(argument.get()))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
} }
ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = ck::utils::conv::get_flops( std::size_t flop = ck::utils::conv::get_flops(
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths); params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
......
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl_bf16_splitk.cpp 0 → 100644
View file @ 6805df0e
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "conv_util.hpp"
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "tensor_layout.hpp"
#include "element_wise_operation.hpp"
#include "device_unary_elementwise.hpp"
#include "device_convnd_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
#include "reference_conv_backward_weight.hpp"
using InDataType = ck::bhalf_t;
using WeiDataType = ck::bhalf_t;
using OutDataType = ck::bhalf_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnaryTypeConvert = ck::tensor_operation::element_wise::UnaryTypeConvert<ck::bhalf_t, float>;
using DeviceUnaryElementwiseTypeConvertInstance = ck::tensor_operation::device::
DeviceUnaryElementwise<AccDataType, WeiDataType, UnaryTypeConvert, 1, 4>;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
using DeviceConvBwdWeightBasePtr =
ck::tensor_operation::device::DeviceConvBwdWeightPtr<InElementOp, WeiElementOp, OutElementOp>;
// clang-format off
template <ck::index_t NumDimSpatial>
using DeviceConvndBwdWeightInstance_bf16_splitk = ck::tensor_operation::device::
DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
InDataType, // InDataType
AccDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
NumDimSpatial, // NumDimSpatial
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
4>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
template <ck::index_t NumDimSpatial>
using ReferenceConvBwdWeightInstance =
ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
template <typename HostTensorB, typename HostTensorA, typename Functor>
void host_elementwise(HostTensorB& B,
const HostTensorA& A,
const std::vector<std::size_t>& shape,
Functor functor)
{
size_t tensor_size = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>{});
std::cout << __LINE__ << ":" << tensor_size << ", " << A.mData[0] << std::endl;
for(std::size_t n = 0; n < tensor_size; ++n)
{
B.mData[n] = functor(A.mData[n]);
}
}
void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: is show log (0=no, 1=yes)\n"
<< "arg5: split-k : in this example split-k must be larger than 1\n"
<< "arg6: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n"
<< " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
<< " <strides>, (ie Sy, Sx for 2D)\n"
<< " <dilations>, (ie Dy, Dx for 2D)\n"
<< " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
<< " <right padding>, (ie RightPy, RightPx for 2D)\n"
<< std::endl;
}
ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
ck::utils::conv::ConvParams params;
int arg_idx = 7;
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
}
DeviceConvBwdWeightBasePtr get_conv_instance(int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 3: {
return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<3>>();
}
case 2: {
return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<2>>();
}
case 1: {
return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<1>>();
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
int do_log = 0;
int split_k = 2;
ck::utils::conv::ConvParams params;
params.C_ = 128;
if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
}
else if(argc > 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
num_dim_spatial = std::stoi(argv[6]);
// check args number
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + 7;
if(cmdline_nargs != argc)
{
print_use_msg();
exit(1);
}
params = parse_conv_params(num_dim_spatial, argv);
}
else if(argc != 1)
{
print_use_msg();
exit(1);
}
if(split_k <= 1)
{
print_use_msg();
exit(1);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
Tensor<InDataType> in_n_c_hi_wi(
ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
Tensor<WeiDataType> wei_k_c_y_x_host_result(
ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<WeiDataType> wei_k_c_y_x_device_result(
ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<OutDataType> out_n_k_ho_wo(
ck::utils::conv::get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_device_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_c_y_x_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
// reset input to zero
wei_device_buf.SetZero();
// do GEMM
auto conv = get_conv_instance(num_dim_spatial);
auto invoker = conv->MakeInvokerPointer();
auto argument =
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{},
split_k);
// alloc work space
size_t bwd_weight_workspace_size = conv->GetWorkSpaceSize(argument.get());
if(bwd_weight_workspace_size <= 0)
{
print_use_msg();
exit(1);
}
float conv_ave_time = 0.f;
DeviceMem wei_work_space_device_buf(bwd_weight_workspace_size);
wei_work_space_device_buf.SetZero();
conv->SetWorkSpacePointer(argument.get(), wei_work_space_device_buf.GetDeviceBuffer());
if(!conv->IsSupportedArgument(argument.get()))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
conv_ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = ck::utils::conv::get_flops(
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / conv_ave_time;
float gb_per_sec = num_btype / 1.E6 / conv_ave_time;
std::cout << "Perf: conv: " << conv_ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << std::endl;
if(do_verification)
{
auto verify_f = [&](const auto& ref_conv) {
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x_host_result,
out_n_k_ho_wo,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
if(do_log)
{
LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
return ck::utils::check_err(wei_k_c_y_x_device_result.mData,
wei_k_c_y_x_host_result.mData)
? 0
: 1;
};
switch(num_dim_spatial)
{
case 3: {
auto ref_conv = ReferenceConvBwdWeightInstance<3>();
verify_f(ref_conv);
break;
}
case 2: {
auto ref_conv = ReferenceConvBwdWeightInstance<2>();
verify_f(ref_conv);
break;
}
case 1: {
auto ref_conv = ReferenceConvBwdWeightInstance<1>();
verify_f(ref_conv);
break;
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
return 0;
}
example/21_gemm_layernorm/CMakeLists.txt
View file @ 6805df0e
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_fp16 gemm_bias_relu_add_layernorm_xdl_fp16.cpp)
add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp) add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp)
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp 0 → 100644
View file @ 6805df0e
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_5ary_elementwise.hpp"
#include "device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
#include "element_wise_reduce_operation.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using ADataType = F16;
using BDataType = F16;
using CDataType = F16;
using C0DataType = F32;
using C1DataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32;
using DDataType = F32;
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
using GammaDataType = F16;
using BetaDataType = F16;
using LayerNormOutDataType = F16;
using NormalizeComputeDataType = F32;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = ck::tensor_operation::element_wise::Relu;
using C1ElementOp = PassThrough;
using ReduceSumOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<ReduceSumOp, ReduceSumOp>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DxsGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmSpecialization =
ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmBiasAddReduceInstance = ck::tensor_operation::device::DeviceGemmBiasAddReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| C1| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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| SrcDstScalarPerVector|
//######| | | | | | | | | | | | | Operation| Operation| 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| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, C1ElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DxsGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
GemmAccDataType,
AElementOp,
BElementOp,
PassThrough>;
using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
using DeviceNormalizeInstance =
ck::tensor_operation::device::Device5AryElementwise<CDataType,
DDataType,
DDataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType,
NormalizeComputeDataType,
NormalizeFunctor,
2,
8,
8, // scalarPerVector: gemm_out
1, // scalarPerVector: reduce_mean
1, // scalarPerVector: reduce_mean_square
8, // scalarPerVector: Gamma
8, // scalarPerVector: Beta
8>; // scalarPerVector: LayerNorm_out
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}),
std::vector<std::size_t>({stride}));
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
template <typename CDataType,
typename DDataType,
typename AccDataType,
typename C0DataType,
typename C1DataType,
typename A_functor,
typename B_functor,
typename C_functor,
typename C1_functor>
void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
const Tensor<ADataType>& a_m_k,
const Tensor<ADataType>& b_k_n,
const Tensor<C0DataType>& bias_n,
const Tensor<C1DataType>& c1_m_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<GammaDataType>& beta_n,
A_functor a_element_op,
B_functor b_element_op,
C_functor c_element_op,
C1_functor c1_element_op,
int M,
int N)
{
int StrideC = N;
Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<DDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
auto averageOpInst = UnaryDivElementOp{N};
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
// c = activation(c + bias) + c1_functor(c1)
for(int m = 0; m < M; ++m)
for(int n = 0; n < N; ++n)
{
AccDataType acc =
static_cast<AccDataType>(c_m_n(m, n)) + static_cast<AccDataType>(bias_n(n));
AccDataType c1 = static_cast<AccDataType>(c1_m_n(m, n));
c_element_op(acc, acc);
c1_element_op(c1, c1);
acc += c1;
c_m_n(m, n) = static_cast<CDataType>(acc);
}
// reduce_mean and reduce_square_mean
auto reduceSumOpInst = ReduceSumOp{};
for(int m = 0; m < M; ++m)
{
auto mean_acc = reduceSumOpInst.GetIdentityValue<AccDataType>();
auto square_mean_acc = reduceSumOpInst.GetIdentityValue<AccDataType>();
for(int n = 0; n < N; ++n)
{
AccDataType c_val = ck::type_convert<AccDataType>(c_m_n(m, n));
AccDataType square_c_val = 0;
UnarySquareElementOp{}(square_c_val, c_val);
reduceSumOpInst(mean_acc, c_val);
reduceSumOpInst(square_mean_acc, square_c_val);
}
averageOpInst(mean_acc, mean_acc);
averageOpInst(square_mean_acc, square_mean_acc);
mean_m(m) = ck::type_convert<DDataType>(mean_acc);
meanSquare_m(m) = ck::type_convert<DDataType>(square_mean_acc);
}
// LayerNorm
auto layerNormInst = NormalizeFunctor{};
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
AccDataType out_acc = 0;
layerNormInst(out_acc,
static_cast<AccDataType>(c_m_n(m, n)),
static_cast<AccDataType>(mean_m(m)),
static_cast<AccDataType>(meanSquare_m(m)),
static_cast<AccDataType>(gamma_n(n)),
static_cast<AccDataType>(beta_n(n)));
out_m_n(m, n) = static_cast<DDataType>(out_acc);
}
}
}
template <typename ADataType,
typename BDataType,
typename CDataType,
typename C0DataType,
typename C1DataType,
typename DDataType,
typename GammaDataType,
typename BetaDataType,
typename NormalizeDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(C0DataType) * M * N +
sizeof(C1DataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
std::size_t normalize_num_byte = sizeof(CDataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M + sizeof(GammaDataType) * N +
sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
float normalize_gb_per_sec = normalize_num_byte / 1.E6 / normalize_time;
std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
<< tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
<< " GB/s, " << std::endl;
}
int main()
{
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideC = 1024;
ck::index_t StrideC1 = 1024;
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<C0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<C1DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> reduceMean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<DDataType> reduceMeanSquare_m(f_host_tensor_descriptor1d(M, 1));
Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
Tensor<LayerNormOutDataType> layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
bias_n.GenerateTensorValue(GeneratorTensor_3<C0DataType>{-1, 1});
c1_m_n.GenerateTensorValue(GeneratorTensor_3<C1DataType>{-5, 5});
gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpace());
DeviceMem bias_device_buf(sizeof(C0DataType) * bias_n.mDesc.GetElementSpace());
DeviceMem c1_device_buf(sizeof(C1DataType) * c1_m_n.mDesc.GetElementSpace());
DeviceMem reduceMean_device_buf(sizeof(DDataType) * reduceMean_m.mDesc.GetElementSpace());
DeviceMem reduceMeanSquare_device_buf(sizeof(DDataType) *
reduceMeanSquare_m.mDesc.GetElementSpace());
DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpace());
DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpace());
DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
layerNorm_m_n.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
bias_device_buf.ToDevice(bias_n.mData.data());
c1_device_buf.ToDevice(c1_m_n.mData.data());
gamma_device_buf.ToDevice(gamma_n.mData.data());
beta_device_buf.ToDevice(beta_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto c1_element_op = C1ElementOp{};
auto dxs_global =
ck::make_tuple(static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()));
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
// Prepare GEMM, reduce_mean, reduce_mean_square
auto gemmReduce = DeviceGemmBiasAddReduceInstance{};
auto gemmReduce_invoker = gemmReduce.MakeInvoker();
auto gemmReduce_argument =
gemmReduce.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<C0DataType*>(bias_device_buf.GetDeviceBuffer()),
static_cast<C1DataType*>(c1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op,
c1_element_op,
dxs_in_element_op,
dxs_out_element_op);
if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
reduceMean_device_buf.SetZero();
reduceMeanSquare_device_buf.SetZero();
// Prepare LayerNorm
auto normalize = DeviceNormalizeInstance{};
auto normalize_invoker = normalize.MakeInvoker();
auto normalize_argument = normalize.MakeArgument(
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()),
static_cast<GammaDataType*>(gamma_device_buf.GetDeviceBuffer()),
static_cast<BetaDataType*>(beta_device_buf.GetDeviceBuffer()),
static_cast<LayerNormOutDataType*>(layerNorm_device_buf.GetDeviceBuffer()),
{M, N},
{StrideC, 1},
{1, 0},
{1, 0},
{0, 1},
{0, 1},
{StrideC, 1},
NormalizeFunctor{});
if(!normalize.IsSupportedArgument(normalize_argument))
{
throw std::runtime_error("The runtime parameters seems not supported by the "
"Device5AryElementwise instance, exiting!");
}
// run kernel
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, false});
bool pass = true;
{
// verification
Tensor<LayerNormOutDataType> host_layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
host_gemm_layernorm<CDataType, DDataType, ReduceAccDataType>(host_layerNorm_m_n,
a_m_k,
b_k_n,
bias_n,
c1_m_n,
gamma_n,
beta_n,
a_element_op,
b_element_op,
c_element_op,
c1_element_op,
M,
N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData,
host_layerNorm_m_n.mData,
"Error: Incorrect results layerNorm_m_n",
1e-2,
1e-2);
}
{
// evaluate kernel perf
bool time_kernel = true;
float gemm_reduce_mean_reduce_square_mean_ave_time =
gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
float normalize_ave_time =
normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, time_kernel});
if(time_kernel)
DumpGemmLayerNormPerf<ADataType,
BDataType,
CDataType,
C0DataType,
C1DataType,
DDataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
}
return pass ? 0 : 1;
}
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
View file @ 6805df0e
...@@ -2,7 +2,6 @@ ...@@ -2,7 +2,6 @@
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include "check_err.hpp" #include "check_err.hpp"
#include "config.hpp" #include "config.hpp"
...@@ -45,17 +44,14 @@ using CLayout = ck::tensor_layout::gemm::RowMajor; ...@@ -45,17 +44,14 @@ using CLayout = ck::tensor_layout::gemm::RowMajor;
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;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ReduceSumOp = ck::reduce::Add<ReduceAccDataType>; using ReduceSumOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<ReduceSumOp, ReduceSumOp>; using DxsReduceOp = ck::Tuple<ReduceSumOp, ReduceSumOp>;
using UnaryIdenticElementOp = using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>; using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnaryDivElementOp = using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>; using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using UnarySquareElementOp = using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DxsGlobalMemOp = using DxsGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd, ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
...@@ -70,7 +66,7 @@ using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_ ...@@ -70,7 +66,7 @@ using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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| SrcDstScalarPerVector| //######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| 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| _MPerBlock| //######| | | | | | | | | | | Operation| 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| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DxsGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; < Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DxsGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
...@@ -143,7 +139,7 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n, ...@@ -143,7 +139,7 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> mean_m(f_host_tensor_descriptor1d(M, 1)); Tensor<DDataType> mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<DDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1)); Tensor<DDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
auto averageOpInst = UnaryDivElementOp{M}; auto averageOpInst = UnaryDivElementOp{N};
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
...@@ -157,13 +153,14 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n, ...@@ -157,13 +153,14 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
auto reduceSumOpInst = ReduceSumOp{}; auto reduceSumOpInst = ReduceSumOp{};
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
float mean_acc = reduceSumOpInst.GetIdentityValue(); auto mean_acc = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
float square_mean_acc = reduceSumOpInst.GetIdentityValue(); auto square_mean_acc = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
ReduceAccDataType c_val = ck::type_convert<float>(c_m_n(m, n)); auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n(m, n));
ReduceAccDataType square_c_val = 0; auto square_c_val = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
UnarySquareElementOp{}(square_c_val, c_val); UnarySquareElementOp{}(square_c_val, c_val);
reduceSumOpInst(mean_acc, c_val); reduceSumOpInst(mean_acc, c_val);
...@@ -183,7 +180,12 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n, ...@@ -183,7 +180,12 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
float out_f32 = 0; float out_f32 = 0;
layerNormInst(out_f32, c_m_n(m, n), mean_m(m), meanSquare_m(m), gamma_n(n), beta_n(n)); layerNormInst(out_f32,
static_cast<float>(c_m_n(m, n)),
static_cast<float>(mean_m(m)),
static_cast<float>(meanSquare_m(m)),
static_cast<float>(gamma_n(n)),
static_cast<float>(beta_n(n)));
out_m_n(m, n) = static_cast<out_type>(out_f32); out_m_n(m, n) = static_cast<out_type>(out_f32);
} }
} }
...@@ -267,8 +269,8 @@ int main() ...@@ -267,8 +269,8 @@ int main()
ck::make_tuple(static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()), ck::make_tuple(static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer())); static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()));
auto dxs_in_element_op = DxsInElementOp{}; auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOp{M, M}; auto dxs_out_element_op = DxsOutElementOps{N, N};
// Prepare GEMM, reduce_mean, reduce_mean_square // Prepare GEMM, reduce_mean, reduce_mean_square
auto gemmReduce = DeviceGemmReduceInstance{}; auto gemmReduce = DeviceGemmReduceInstance{};
......
include/ck/tensor_operation/gpu/device/device_5ary_elementwise.hpp
View file @ 6805df0e
...@@ -3,7 +3,6 @@ ...@@ -3,7 +3,6 @@
#include <sstream> #include <sstream>
#include "device.hpp" #include "device.hpp"
#include "device_base.hpp" #include "device_base.hpp"
#include "common_header.hpp"
#include "gridwise_5ary_Elementwise_1d.hpp" #include "gridwise_5ary_Elementwise_1d.hpp"
#include "tensor_layout.hpp" #include "tensor_layout.hpp"
#include "tensor_descriptor.hpp" #include "tensor_descriptor.hpp"
......
include/ck/tensor_operation/gpu/device/device_base.hpp
View file @ 6805df0e
...@@ -15,6 +15,8 @@ struct BaseArgument ...@@ -15,6 +15,8 @@ struct BaseArgument
BaseArgument& operator=(const BaseArgument&) = default; BaseArgument& operator=(const BaseArgument&) = default;
virtual ~BaseArgument() {} virtual ~BaseArgument() {}
void* p_workspace_ = nullptr;
}; };
struct BaseInvoker struct BaseInvoker
...@@ -42,7 +44,11 @@ struct BaseOperator ...@@ -42,7 +44,11 @@ struct BaseOperator
virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; } virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; }
virtual void SetWorkSpacePointer(BaseArgument*, void*) const {} virtual void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const
{
assert(p_arg);
p_arg->p_workspace_ = p_workspace;
}
virtual ~BaseOperator() {} virtual ~BaseOperator() {}
}; };
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp
View file @ 6805df0e
...@@ -22,7 +22,7 @@ template <typename GridwiseGemm, ...@@ -22,7 +22,7 @@ template <typename GridwiseGemm,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation, typename CElementwiseOperation,
typename DxsInElementwiseOperation, typename DxsInElementwiseOperation,
typename DxsAccElementwiseOperation, typename DxsReduceAccElementwiseOperation,
typename AGridDesc_AK0_M_AK1, typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1, typename BGridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
...@@ -44,7 +44,7 @@ __global__ void ...@@ -44,7 +44,7 @@ __global__ void
const BElementwiseOperation b_element_op, const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op, const CElementwiseOperation c_element_op,
const DxsInElementwiseOperation dxs_in_element_op, const DxsInElementwiseOperation dxs_in_element_op,
const DxsAccElementwiseOperation dxs_out_element_op, const DxsReduceAccElementwiseOperation dxs_out_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1, const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1, const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
...@@ -126,7 +126,7 @@ template <typename ALayout, ...@@ -126,7 +126,7 @@ template <typename ALayout,
typename CElementwiseOperation, typename CElementwiseOperation,
typename DxsReduceOperation, typename DxsReduceOperation,
typename DxsInElementwiseOperation, typename DxsInElementwiseOperation,
typename DxsAccElementwiseOperation, typename DxsReduceAccElementwiseOperation,
typename DGlobalMemoryDataOperation, typename DGlobalMemoryDataOperation,
GemmSpecialization GemmSpec, GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage, index_t NumGemmKPrefetchStage,
...@@ -162,12 +162,12 @@ template <typename ALayout, ...@@ -162,12 +162,12 @@ template <typename ALayout,
index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock, index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock, index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()> LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal, struct DeviceBatchedGemmReduce_Xdl_CShuffle
AElementwiseOperation, : public DeviceGemmReduce<AElementwiseOperation,
BElementwiseOperation, BElementwiseOperation,
CElementwiseOperation, CElementwiseOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsAccElementwiseOperation> DxsReduceAccElementwiseOperation>
{ {
using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle; using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle;
...@@ -527,7 +527,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -527,7 +527,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
CElementwiseOperation, CElementwiseOperation,
DxsReduceOperation, DxsReduceOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsAccElementwiseOperation, DxsReduceAccElementwiseOperation,
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
DGlobalMemoryDataOperation, DGlobalMemoryDataOperation,
AGridDesc_AK0_M_AK1, AGridDesc_AK0_M_AK1,
...@@ -587,7 +587,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -587,7 +587,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op, CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op, DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op, DxsReduceAccElementwiseOperation dxs_out_element_op,
index_t BatchCount) index_t BatchCount)
: p_a_grid_{p_a_grid}, : p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid}, p_b_grid_{p_b_grid},
...@@ -645,7 +645,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -645,7 +645,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op_; BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_; CElementwiseOperation c_element_op_;
DxsInElementwiseOperation dxs_in_element_op_; DxsInElementwiseOperation dxs_in_element_op_;
DxsAccElementwiseOperation dxs_out_element_op_; DxsReduceAccElementwiseOperation dxs_out_element_op_;
}; };
// Invoker // Invoker
...@@ -703,7 +703,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -703,7 +703,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation, BElementwiseOperation,
CElementwiseOperation, CElementwiseOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsAccElementwiseOperation, DxsReduceAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1, DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1, DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
...@@ -746,7 +746,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -746,7 +746,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation, BElementwiseOperation,
CElementwiseOperation, CElementwiseOperation,
DxsInElementwiseOperation, DxsInElementwiseOperation,
DxsAccElementwiseOperation, DxsReduceAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1, DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1, DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
...@@ -832,7 +832,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -832,7 +832,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op, CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op, DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op, DxsReduceAccElementwiseOperation dxs_out_element_op,
index_t BatchCount) index_t BatchCount)
{ {
return Argument{p_a, return Argument{p_a,
...@@ -856,27 +856,29 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba ...@@ -856,27 +856,29 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
static auto MakeInvoker() { return Invoker{}; } static auto MakeInvoker() { return Invoker{}; }
// polymorphic // polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a, std::unique_ptr<BaseArgument>
const void* p_b, MakeArgumentPointer(const void* p_a,
void* p_c, const void* p_b,
DPtrsGlobal p_dxs, void* p_c,
index_t MRaw, void* p_dxs,
index_t NRaw, index_t MRaw,
index_t KRaw, index_t NRaw,
index_t StrideA, index_t KRaw,
index_t StrideB, index_t StrideA,
index_t StrideC, index_t StrideB,
AElementwiseOperation a_element_op, index_t StrideC,
BElementwiseOperation b_element_op, AElementwiseOperation a_element_op,
CElementwiseOperation c_element_op, BElementwiseOperation b_element_op,
DxsInElementwiseOperation dxs_in_element_op, CElementwiseOperation c_element_op,
DxsAccElementwiseOperation dxs_out_element_op, DxsInElementwiseOperation dxs_in_element_op,
index_t BatchCount) override DxsReduceAccElementwiseOperation dxs_out_element_op,
index_t BatchCount) override
{ {
DPtrsGlobal dxs_tuple = *(static_cast<DPtrsGlobal*>(p_dxs));
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a), return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b), static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c), static_cast<CDataType*>(p_c),
p_dxs, dxs_tuple,
MRaw, MRaw,
NRaw, NRaw,
KRaw, KRaw,
......
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