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
Expand all 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() {
echo "NODE_NAME = \$NODE_NAME"
lsb_release -sd
uname -r
cat /sys/module/amdgpu/version
ls /opt/ -la
"""
}
......
example/12_reduce/reduce_blockwise.cpp
View file @ 6805df0e
......@@ -33,11 +33,11 @@ constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
using DeviceReduceInstance = DeviceReduceMultiBlock<InDataType,
AccDataType,
......@@ -247,6 +247,13 @@ int main(int argc, char* argv[])
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)
{
ReductionHost<InDataType,
......@@ -261,8 +268,13 @@ int main(int argc, char* argv[])
OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
hostReduce.Run(alpha,
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;
......@@ -277,20 +289,19 @@ int main(int argc, char* argv[])
auto reduce = DeviceReduceInstance{};
auto argument_ptr = reduce.MakeArgumentPointer(
i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
out_index_dev.GetDeviceBuffer(),
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)},
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
auto argument_ptr = reduce.MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
out_index_dev.GetDeviceBuffer(),
in_elementwise_op,
acc_elementwise_op);
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;
constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
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,
AccDataType,
......@@ -184,6 +184,13 @@ int main(int argc, char* argv[])
if(beta != 0.0f)
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)
{
ReductionHost<InOutDataType,
......@@ -198,7 +205,13 @@ int main(int argc, char* argv[])
OutputIndex>
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;
......@@ -217,20 +230,19 @@ int main(int argc, char* argv[])
auto reduce_1 = DeviceReduceInstance_1{};
auto argument_ptr_1 = reduce_1.MakeArgumentPointer(
i_inLengths_1,
i_inStrides_1,
i_inLengths_2,
i_inStrides_2,
reduceDims_1,
1.0f,
0.0f,
in_1_dev.GetDeviceBuffer(),
nullptr,
in_2_dev.GetDeviceBuffer(),
nullptr,
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)},
PassThroughOp{});
auto argument_ptr_1 = reduce_1.MakeArgumentPointer(i_inLengths_1,
i_inStrides_1,
i_inLengths_2,
i_inStrides_2,
reduceDims_1,
1.0f,
0.0f,
in_1_dev.GetDeviceBuffer(),
nullptr,
in_2_dev.GetDeviceBuffer(),
nullptr,
in_elementwise_op,
PassThroughOp{});
if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
{
......@@ -243,20 +255,19 @@ int main(int argc, char* argv[])
auto reduce_2 = DeviceReduceInstance_2{};
auto argument_ptr_2 = reduce_2.MakeArgumentPointer(
i_inLengths_2,
i_inStrides_2,
i_outLengths,
i_outStrides,
reduceDims_2,
alpha,
beta,
in_2_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
nullptr,
PassThroughOp{},
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
auto argument_ptr_2 = reduce_2.MakeArgumentPointer(i_inLengths_2,
i_inStrides_2,
i_outLengths,
i_outStrides,
reduceDims_2,
alpha,
beta,
in_2_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
nullptr,
PassThroughOp{},
acc_elementwise_op);
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,
const std::array<ck::index_t, 2>& in_left_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 InElementwiseOperation = typename ck::
reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation = typename ck::
reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
const InElementwiseOperation in_elementwise_op(divider);
const AccElementwiseOperation acc_elementwise_op(divider);
auto elementwise_ops =
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)
{
......@@ -48,7 +47,7 @@ static void pool_host_verify(const Tensor<InDataType>& in,
ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
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)
{
......@@ -86,7 +85,7 @@ static void pool_host_verify(const Tensor<InDataType>& in,
AccDataType,
IndexDataType>;
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOperation::GetIdentityValue();
auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
IndexDataType accuIndex = 0;
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;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using DsReduceOp = ck::Tuple<ck::reduce::Max<ReduceAccDataType>>;
using DsElementOp = ck::Tuple<
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>>;
using DsReduceOp = ck::Tuple<ck::reduce::Max>;
using DsElementOp = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
using DGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
......@@ -236,10 +235,14 @@ int main(int argc, char* argv[])
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)
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;
}
......
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;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<ReduceAccDataType>;
using D1ReduceOp = ck::reduce::Add<ReduceAccDataType>;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
using UnaryDivElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
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 DGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
......@@ -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|
//######| | | | | | | | | | | 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
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
......@@ -204,8 +201,8 @@ int main(int argc, char* argv[])
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
auto dxs_in_element_op = DxsInElementOp{};
auto dxs_out_element_op = DxsOutElementOp{M, M};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
......@@ -261,14 +258,14 @@ int main(int argc, char* argv[])
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d0_val = 0;
float d1_val = 0;
auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_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;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<ReduceAccDataType>;
using D1ReduceOp = ck::reduce::Add<ReduceAccDataType>;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using DGlobalMemOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
......@@ -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|
//######| | | | | | | | | | | 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
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
......@@ -206,8 +204,8 @@ int main(int argc, char* argv[])
a_element_op,
b_element_op,
c_element_op,
DxsInElementOp{},
DxsOutElementOp{},
DxsInElementOps{},
DxsOutElementOps{},
BatchCount);
if(!batched_gemm.IsSupportedArgument(argument))
......@@ -259,14 +257,15 @@ int main(int argc, char* argv[])
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
float c_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d0_val = 0;
float d1_val = 0;
auto c_val =
ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
UnaryIdenticElementOp{}(d0_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;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
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;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
......
example/19_binary_elementwise/elementwise_add_1d.cpp
View file @ 6805df0e
......@@ -42,8 +42,7 @@ using ABDataType = F16;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
......
example/19_binary_elementwise/elementwise_add_4d.cpp
View file @ 6805df0e
......@@ -42,8 +42,7 @@ using ABDataType = F16;
using CDataType = F16;
using EltwiseComputeDataType = F32;
using Add = ck::tensor_operation::binary_element_wise::
Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
using Add = ck::tensor_operation::element_wise::Add;
using DeviceElementwiseAddInstance =
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)
target_link_libraries(example_convnd_bwd_weight_xdl PRIVATE conv_util)
\ No newline at end of file
add_example_executable(example_convnd_bwd_weight_xdl_bf16_splitk convnd_bwd_weight_xdl_bf16_splitk.cpp)
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[])
split_k);
// alloc work space
size_t bwd_weight_workspace_size = conv->GetWorkSpaceSize(argument.get());
float ave_time = 0.f;
if(std::is_same<InDataType, ck::bhalf_t>::value && split_k > 1)
float ave_time = 0.f;
if(!conv->IsSupportedArgument(argument.get()))
{
DeviceMem wei_work_space_device_buf(bwd_weight_workspace_size);
wei_work_space_device_buf.SetZero();
argument = conv->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
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});
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});
std::size_t flop = ck::utils::conv::get_flops(
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)
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
View file @ 6805df0e
......@@ -2,7 +2,6 @@
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "check_err.hpp"
#include "config.hpp"
......@@ -45,17 +44,14 @@ using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = 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 UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
using UnaryDivElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
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,
......@@ -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|
//######| | | | | | | | | | | 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
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
......@@ -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<DDataType> mean_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_invoker = ref_gemm.MakeInvoker();
......@@ -157,13 +153,14 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
auto reduceSumOpInst = ReduceSumOp{};
for(int m = 0; m < M; ++m)
{
float mean_acc = reduceSumOpInst.GetIdentityValue();
float square_mean_acc = reduceSumOpInst.GetIdentityValue();
auto mean_acc = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
auto square_mean_acc = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType c_val = ck::type_convert<float>(c_m_n(m, n));
ReduceAccDataType square_c_val = 0;
auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n(m, n));
auto square_c_val = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
UnarySquareElementOp{}(square_c_val, c_val);
reduceSumOpInst(mean_acc, c_val);
......@@ -183,7 +180,12 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
for(int n = 0; n < N; ++n)
{
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);
}
}
......@@ -267,8 +269,8 @@ int main()
ck::make_tuple(static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()));
auto dxs_in_element_op = DxsInElementOp{};
auto dxs_out_element_op = DxsOutElementOp{M, M};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
// Prepare GEMM, reduce_mean, reduce_mean_square
auto gemmReduce = DeviceGemmReduceInstance{};
......
include/ck/tensor_operation/gpu/device/device_5ary_elementwise.hpp
View file @ 6805df0e
......@@ -3,7 +3,6 @@
#include <sstream>
#include "device.hpp"
#include "device_base.hpp"
#include "common_header.hpp"
#include "gridwise_5ary_Elementwise_1d.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
......
include/ck/tensor_operation/gpu/device/device_base.hpp
View file @ 6805df0e
......@@ -15,6 +15,8 @@ struct BaseArgument
BaseArgument& operator=(const BaseArgument&) = default;
virtual ~BaseArgument() {}
void* p_workspace_ = nullptr;
};
struct BaseInvoker
......@@ -42,7 +44,11 @@ struct BaseOperator
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() {}
};
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp
View file @ 6805df0e
......@@ -22,7 +22,7 @@ template <typename GridwiseGemm,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsAccElementwiseOperation,
typename DxsReduceAccElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
......@@ -44,7 +44,7 @@ __global__ void
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_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 BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
......@@ -126,7 +126,7 @@ template <typename ALayout,
typename CElementwiseOperation,
typename DxsReduceOperation,
typename DxsInElementwiseOperation,
typename DxsAccElementwiseOperation,
typename DxsReduceAccElementwiseOperation,
typename DGlobalMemoryDataOperation,
GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
......@@ -162,12 +162,12 @@ template <typename ALayout,
index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation>
struct DeviceBatchedGemmReduce_Xdl_CShuffle
: public DeviceGemmReduce<AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsReduceAccElementwiseOperation>
{
using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle;
......@@ -527,7 +527,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
CElementwiseOperation,
DxsReduceOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation,
DxsReduceAccElementwiseOperation,
InMemoryDataOperationEnum::Set,
DGlobalMemoryDataOperation,
AGridDesc_AK0_M_AK1,
......@@ -587,7 +587,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op,
DxsReduceAccElementwiseOperation dxs_out_element_op,
index_t BatchCount)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
......@@ -645,7 +645,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
DxsInElementwiseOperation dxs_in_element_op_;
DxsAccElementwiseOperation dxs_out_element_op_;
DxsReduceAccElementwiseOperation dxs_out_element_op_;
};
// Invoker
......@@ -703,7 +703,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation,
DxsReduceAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
......@@ -746,7 +746,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsAccElementwiseOperation,
DxsReduceAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
......@@ -832,7 +832,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op,
DxsReduceAccElementwiseOperation dxs_out_element_op,
index_t BatchCount)
{
return Argument{p_a,
......@@ -856,27 +856,29 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGloba
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
DPtrsGlobal p_dxs,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsAccElementwiseOperation dxs_out_element_op,
index_t BatchCount) override
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
void* p_dxs,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
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),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
p_dxs,
dxs_tuple,
MRaw,
NRaw,
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