Skip to content

GitLab

Unverified Commit 6c3c06bf authored by rocking5566's avatar rocking5566 Committed by GitHub
Browse files

Gemm multiple d multiple r (#335) 

* Imitate XXX_gemm_multiple_d, add XXX_gemm_multiple_d_multiple_r for gemm + reduction

* Implement run of kernel

* Add example

* Fix parameter of typo

* Rewrite the reduceMax example

* Rewrite the reduceMean + reduceMeanSquare example

* Refine naming

* Refine folder name

* refine naming

* Rewrite the gemm + bias + relu + add + layernorm example

* Rewrite the gemm + layernorm example

* clang-format

* Fix bug if sync lds

* Fix compile error
parent cac014f1
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2950 additions and 912 deletions
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt 0 → 100644
View file @ 6c3c06bf
add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp 0 → 100644
View file @ 6c3c06bf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.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;
// DataType
using ADataType = F16;
using BDataType = F16;
using GemmAccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F16;
using D1DataType = F16;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = F16;
using ReduceAccDataType = F32;
using R0DataType = F32;
using R1DataType = F32;
using RsDataType = ck::Tuple<R0DataType, R1DataType>;
// Layout
using ALayout = Row;
using BLayout = Col;
using D1Layout = Row;
using ELayout = D1Layout;
// Elementwise op
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAdd = ck::tensor_operation::element_wise::AddAdd;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAdd;
using QsElementOp = ck::Tuple<PassThrough, Square>;
using RsElementOp = ck::Tuple<Div, Div>;
// ReduceOp
using R0ThreadReduceOp = ck::reduce::Add;
using R1ThreadReduceOp = ck::reduce::Add;
using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
//######| ALayout| BLayout| ELayout| AData| BData| GemmAccData| CShuffle| DsData| EData| ReduceAccData| RsData| A| B| CDE| Qs| Rs| Thread| Global| 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| CDRThreadTransfer| CDE| RThreadTransfer|
//######| | | | Type| Type| Type| DataType| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
//######| | | | | | | | | | | | Operation| 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| _MPerBlock_NPerBlock| ScalarPerVector| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NPerBlock| |
< ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType, AElementOp, BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
EDataType,
GemmAccDataType,
AElementOp,
BElementOp,
PassThrough>;
template <typename ADataType,
typename BDataType,
typename D0DataType,
typename D1DataType,
typename EDataType,
typename R0DataType,
typename R1DataType>
void DumpPerf(float ave_time, int M, int N, int K)
{
std::size_t 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(D0DataType) * M * N + sizeof(D1DataType) * M * N +
sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
sizeof(R1DataType) * M;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
<< " GB/s, " << std::endl;
}
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}));
}
};
int main()
{
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 StrideD0 = 0;
ck::index_t StrideD1 = 1024;
ck::index_t StrideE = 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<D0DataType> d0_n(f_host_tensor_descriptor1d(N, 1));
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
d0_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_n.mDesc.GetElementSpaceSize());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_device_buf.ToDevice(d0_n.mData.data());
d1_device_buf.ToDevice(d1_m_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto qs_element_op = QsElementOp{};
auto rs_element_op = RsElementOp{N, N};
// Prepare GEMM, mean, mean_square
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
{r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
M,
N,
K,
StrideA,
StrideB,
{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op,
qs_element_op,
rs_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
// init reducetion buffer to 0
r0_device_buf.SetZero();
r1_device_buf.SetZero();
invoker.Run(argument, StreamConfig{nullptr, false});
bool do_verification = true;
bool pass = true;
if(do_verification)
{
auto I0 = ck::Number<0>{};
auto I1 = ck::Number<1>{};
Tensor<EDataType> e_m_n_host(e_m_n.mDesc);
Tensor<R0DataType> r0_m_host(r0_m.mDesc);
Tensor<R1DataType> r1_m_host(r1_m.mDesc);
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
auto reduce0_op = R0ThreadReduceOp{};
auto reduce1_op = R1ThreadReduceOp{};
for(int m = 0; m < M; ++m)
{
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType square_e_val;
auto e_val = ck::type_convert<GemmAccDataType>(e_m_n_host(m, n));
auto d0_val = ck::type_convert<GemmAccDataType>(d0_n(n));
auto d1_val = ck::type_convert<GemmAccDataType>(d1_m_n(m, n));
cde_element_op(e_val, e_val, d0_val, d1_val);
e_m_n_host(m, n) = ck::type_convert<EDataType>(e_val);
auto e_val_reduce = ck::type_convert<ReduceAccDataType>(e_val);
qs_element_op[I1](square_e_val, e_val_reduce);
reduce0_op(reduce0_acc, e_val_reduce);
reduce1_op(reduce1_acc, square_e_val);
}
rs_element_op[I0](reduce0_acc, reduce0_acc);
rs_element_op[I1](reduce1_acc, reduce1_acc);
r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
}
e_device_buf.FromDevice(e_m_n.mData.data());
r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data());
pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host.mData, "Error: Incorrect results c", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
}
bool time_kernel = true;
if(time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
DumpPerf<ADataType, BDataType, D0DataType, D1DataType, EDataType, R0DataType, R1DataType>(
ave_time, M, N, K);
}
return pass ? 0 : 1;
}
example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp 0 → 100644
View file @ 6c3c06bf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using F64 = double;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
// DataType
using ADataType = F16;
using BDataType = F16;
using GemmAccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ReduceAccDataType = F32;
using R0DataType = F32;
using RsDataType = ck::Tuple<R0DataType>;
// Layout
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
// Elementwise op
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
using QsElementOp = ck::Tuple<PassThrough>;
using RsElementOp = ck::Tuple<PassThrough>;
// ReduceOp
using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
using RsGlobalReduceOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
//######| ALayout| BLayout| ELayout| AData| BData| GemmAccData| CShuffle| DsData| EData| ReduceAccData| RsData| A| B| CDE| Qs| Rs| Thread| Global| 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| CDRThreadTransfer| CDE| RThreadTransfer|
//######| | | | Type| Type| Type| DataType| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
//######| | | | | | | | | | | | Operation| 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| _MPerBlock_NPerBlock| ScalarPerVector| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NPerBlock| |
< ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType, AElementOp, BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
EDataType,
GemmAccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
template <typename ADataType, typename BDataType, typename EDataType, typename R0DataType>
void DumpPerf(float ave_time, int M, int N, int K)
{
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(EDataType) * M * N + sizeof(R0DataType) * M;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
<< " GB/s, " << std::endl;
}
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}));
}
};
int main()
{
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 StrideE = 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<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto qs_element_op = QsElementOp{};
auto rs_element_op = RsElementOp{};
// Prepare GEMM, max
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
{r0_device_buf.GetDeviceBuffer()},
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op,
qs_element_op,
rs_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
// [CAUSION]: launch_and_time_kernel will not initialize D.
// If we evaluate kernel multiple time but without initialize D. Verification will fail
r0_device_buf.SetValue(ck::NumericLimits<R0DataType>::Lowest());
invoker.Run(argument, StreamConfig{nullptr, false});
bool do_verification = true;
bool pass = true;
if(do_verification)
{
auto I0 = ck::Number<0>{};
Tensor<EDataType> e_m_n_host(e_m_n.mDesc);
Tensor<R0DataType> r0_m_host(r0_m.mDesc);
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, cde_element_op);
ref_invoker.Run(ref_argument);
auto reduce0_op = RsThreadReduceOp{}[I0];
for(int m = 0; m < M; ++m)
{
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n_host(m, n));
reduce0_op(reduce0_acc, e_val);
};
r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
}
e_device_buf.FromDevice(e_m_n.mData.data());
r0_device_buf.FromDevice(r0_m.mData.data());
pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host.mData, "Error: Incorrect results c", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2);
}
bool time_kernel = true;
if(time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
DumpPerf<ADataType, BDataType, EDataType, R0DataType>(ave_time, M, N, K);
}
return pass ? 0 : 1;
}
example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp 0 → 100644
View file @ 6c3c06bf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.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;
// DataType
using ADataType = F16;
using BDataType = F16;
using GemmAccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ReduceAccDataType = F32;
using R0DataType = F32;
using R1DataType = F32;
using RsDataType = ck::Tuple<R0DataType, R1DataType>;
// Layout
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
// Elementwise op
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
using QsElementOp = ck::Tuple<PassThrough, Square>;
using RsElementOp = ck::Tuple<Div, Div>;
// ReduceOp
using R0ThreadReduceOp = ck::reduce::Add;
using R1ThreadReduceOp = ck::reduce::Add;
using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
//######| ALayout| BLayout| ELayout| AData| BData| GemmAccData| CShuffle| DsData| EData| ReduceAccData| RsData| A| B| CDE| Qs| Rs| Thread| Global| 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| CDRThreadTransfer| CDE| RThreadTransfer|
//######| | | | Type| Type| Type| DataType| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
//######| | | | | | | | | | | | Operation| 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| _MPerBlock_NPerBlock| ScalarPerVector| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NPerBlock| |
< ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType, AElementOp, BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
EDataType,
GemmAccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
template <typename ADataType,
typename BDataType,
typename EDataType,
typename R0DataType,
typename R1DataType>
void DumpPerf(float ave_time, int M, int N, int K)
{
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
sizeof(R1DataType) * M;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
<< " GB/s, " << std::endl;
}
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}));
}
};
int main()
{
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 StrideE = 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<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto qs_element_op = QsElementOp{};
auto rs_element_op = RsElementOp{N, N};
// Prepare GEMM, mean, mean_square
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
{r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op,
qs_element_op,
rs_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
// init reducetion buffer to 0
r0_device_buf.SetZero();
r1_device_buf.SetZero();
invoker.Run(argument, StreamConfig{nullptr, false});
bool do_verification = true;
bool pass = true;
if(do_verification)
{
auto I0 = ck::Number<0>{};
auto I1 = ck::Number<1>{};
Tensor<EDataType> e_m_n_host(e_m_n.mDesc);
Tensor<R0DataType> r0_m_host(r0_m.mDesc);
Tensor<R1DataType> r1_m_host(r1_m.mDesc);
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
auto reduce0_op = R0ThreadReduceOp{};
auto reduce1_op = R1ThreadReduceOp{};
for(int m = 0; m < M; ++m)
{
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType square_e_val;
auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n_host(m, n));
qs_element_op[I1](square_e_val, e_val);
reduce0_op(reduce0_acc, e_val);
reduce1_op(reduce1_acc, square_e_val);
}
rs_element_op[I0](reduce0_acc, reduce0_acc);
rs_element_op[I1](reduce1_acc, reduce1_acc);
r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
}
e_device_buf.FromDevice(e_m_n.mData.data());
r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data());
pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host.mData, "Error: Incorrect results c", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
}
bool time_kernel = true;
if(time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
DumpPerf<ADataType, BDataType, EDataType, R0DataType, R1DataType>(ave_time, M, N, K);
}
return pass ? 0 : 1;
}
example/16_gemm_reduce/CMakeLists.txt deleted 100644 → 0
View file @ cac014f1
add_example_executable(example_gemm_reduce_xdl_max_fp16 gemm_reduce_xdl_max_fp16.cpp)
add_example_executable(example_gemm_reduce_xdl_mean_squaremean_fp16 gemm_reduce_xdl_mean_squaremean_fp16.cpp)
example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp deleted 100644 → 0
View file @ cac014f1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using F64 = double;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using ADataType = F16;
using BDataType = F16;
using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32;
using ReduceDataType = F64;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*>;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
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 ReduceOps = ck::Tuple<ck::reduce::Max>;
using ReduceElementOps = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
static constexpr auto GemmSpecialization =
ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| ReduceData| A| B| C| Reduce| ReduceInEleOp| ReduceAccEleOp| Reduce| 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| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Operation| | | 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| | 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, ReduceAccDataType, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, ReduceOps, ReduceElementOps, ReduceElementOps, ReduceGlobalMemOps, 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,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename ReduceDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M;
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;
std::cout << "gemm + reduceMax Perf: " << gemm_reduce_time << " ms, " << tflops << " TFlops, "
<< gemm_gb_per_sec << " GB/s, " << std::endl;
}
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 1)
{
// do nothing
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](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}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> reduce_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> reduce_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "reduce_m: " << reduce_m_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
DeviceMem reduce_device_buf(sizeof(ReduceDataType) *
reduce_m_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto reduce_element_op = ReduceElementOps{}[ck::Number<0>{}];
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
std::array<void*, 1> reduce_element_ops = {&reduce_element_op};
std::array<void*, 1> p_reduces = {reduce_device_buf.GetDeviceBuffer()};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{},
gemm_element_ops,
{},
reduce_element_ops,
reduce_element_ops);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
// [CAUSION]: launch_and_time_kernel will not initialize D.
// If we evaluate kernel multiple time but without initialize D. Verification will fail
reduce_device_buf.SetValue(ck::NumericLimits<ReduceDataType>::Lowest());
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
reduce_device_buf.FromDevice(reduce_m_device_result.mData.data());
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_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
auto reduce_op = ReduceOps{}[ck::Number<0>{}];
for(int m = 0; m < M; ++m)
{
ReduceAccDataType reduce_acc = reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType curr_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
reduce_op(reduce_acc, curr_val);
};
reduce_m_host_result(m) = reduce_acc;
}
pass = ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"Error: Incorrect results c") &&
ck::utils::check_err(reduce_m_device_result.mData,
reduce_m_host_result.mData,
"Error: Incorrect results d",
1e-3,
1e-3);
}
if(time_kernel)
{
float gemm_reduceMax_ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, ReduceDataType>(
gemm_reduceMax_ave_time, M, N, K);
}
return pass ? 0 : 1;
}
example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_fp16.cpp deleted 100644 → 0
View file @ cac014f1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.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 GemmAccDataType = F32;
using ReduceAccDataType = F32;
using ReduceDataType = F32;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*, ReduceDataType*>;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
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 ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceOp0, ReduceOp1>;
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 ReduceInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using ReduceOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmSpecialization =
ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| ReduceDData| A| B| C| Reduce| ReduceInEleOp| ReduceOutEleOp| Reduce| 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| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Operation| | | 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| | 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, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceGlobalMemOps, 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,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename ReduceDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M +
sizeof(ReduceDataType) * M;
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;
std::cout << "gemm + reduce_mean + reduce_mean_square Perf: " << gemm_reduce_time << " ms, "
<< tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
}
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 1)
{
// do nothing
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](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}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> reduce0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<ReduceDataType> reduce1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> reduce0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<ReduceDataType> reduce1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
reduce0_m_device_result.mDesc.GetElementSpaceSize());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
reduce1_m_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
auto div = UnaryDivElementOp{N};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&div, &div};
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{},
gemm_element_ops,
{},
reduce_in_element_ops,
reduce_out_element_ops);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
// init reducetion buffer to 0
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
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_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
auto reduce0_op = ReduceOp0{};
auto reduce1_op = ReduceOp1{};
for(int m = 0; m < M; ++m)
{
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType square_c_val;
square(square_c_val, c_val);
reduce0_op(reduce0_acc, c_val);
reduce1_op(reduce1_acc, square_c_val);
}
div(reduce0_acc, reduce0_acc);
div(reduce1_acc, reduce1_acc);
reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
pass = ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"Error: Incorrect results c") &&
ck::utils::check_err(reduce0_m_device_result.mData,
reduce0_m_host_result.mData,
"Error: Incorrect results d0",
1e-4,
1e-5) &&
ck::utils::check_err(reduce1_m_device_result.mData,
reduce1_m_host_result.mData,
"Error: Incorrect results d1",
1e-3,
1e-5);
}
if(time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, ReduceDataType>(ave_time, M, N, K);
}
return pass ? 0 : 1;
}
example/CMakeLists.txt
View file @ 6c3c06bf
......@@ -30,7 +30,7 @@ add_subdirectory(12_reduce)
add_subdirectory(13_pool2d_fwd)
add_subdirectory(14_gemm_xdl_requant_relu_requant)
add_subdirectory(15_grouped_gemm)
add_subdirectory(16_gemm_reduce)
add_subdirectory(16_gemm_multi_d_multi_reduces)
add_subdirectory(17_convnd_bwd_data)
add_subdirectory(18_batched_gemm_reduce)
add_subdirectory(19_binary_elementwise)
......
include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp 0 → 100644
View file @ 6c3c06bf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// FIXME: DeviceGemmReduce type need to well define the problem
template <typename ALayout,
typename BLayout,
typename DELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename RsDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename QsElementwiseOperation,
typename RsElementwiseOperation>
struct DeviceGemmMultipleDMultipleR : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr index_t NumRTensor = RsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
std::array<void*, NumRTensor> p_rs,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
std::array<ck::index_t, NumDTensor> StrideDs,
ck::index_t StrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
QsElementwiseOperation qs_element_op,
RsElementwiseOperation rs_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename ALayout,
typename BLayout,
typename DELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename RsDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename QsElementwiseOperation,
typename RsElementwiseOperation>
using DeviceGemmMultipleDMultipleRPtr =
std::unique_ptr<DeviceGemmMultipleDMultipleR<ALayout,
BLayout,
DELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
RsDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
QsElementwiseOperation,
RsElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
View file @ 6c3c06bf
......@@ -130,6 +130,35 @@ struct AddHardswishAdd
}
};
// C = A * B
// E = C + D0 + D1
struct AddAdd
{
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const
{
// Only support floating so far
static_assert(is_same<E, half_t>::value || is_same<E, float>::value ||
is_same<E, double>::value,
"Data type is not supported by this operation!");
static_assert(is_same<C, half_t>::value || is_same<C, float>::value ||
is_same<C, double>::value,
"Data type is not supported by this operation!");
static_assert(is_same<D0, half_t>::value || is_same<D0, float>::value ||
is_same<D0, double>::value,
"Data type is not supported by this operation!");
static_assert(is_same<D1, half_t>::value || is_same<D1, float>::value ||
is_same<D1, double>::value,
"Data type is not supported by this operation!");
const C y = c + type_convert<C>(d0) + type_convert<C>(d1);
e = type_convert<E>(y);
}
};
// C = A * B
// E = FastGelu(C + D0 + D1)
struct AddAddFastGelu
......
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