Skip to content

GitLab

"magic_pdf/git@developer.sourcefind.cn:wangsen/mineru.git" did not exist on "9d6be7c95b4853fe000c2020c261bb941ba53e03"
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
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/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp
View file @ 6c3c06bf
......@@ -9,7 +9,7 @@
#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_bias_add_reduce_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/device_5ary_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......@@ -28,57 +28,64 @@ 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 CDataType = F16;
using BiasDataType = F32;
using D0DataType = 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 ReduceDataType = F32;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*, ReduceDataType*>;
using R0DataType = F32;
using R1DataType = F32;
using RsDataType = ck::Tuple<R0DataType, R1DataType>;
using GammaDataType = F16;
using BetaDataType = F16;
using LayerNormOutDataType = F16;
using NormalizeComputeDataType = F32;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = ck::tensor_operation::element_wise::Relu;
using D0ElementOp = PassThrough;
using ReduceSumOp = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceSumOp, ReduceSumOp>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using 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;
// Layout
using ALayout = Row;
using BLayout = Col;
using D1Layout = Row;
using ELayout = D1Layout;
// Elementwise op
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddReluAdd = ck::tensor_operation::element_wise::AddReluAdd;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddReluAdd;
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 DeviceGemmBiasAddReduceInstance = ck::tensor_operation::device::DeviceGemmBiasAddReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data| GemmAcc| CShuffle| ReduceAcc| ReduceData| A| B| C| C1| 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| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| 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| | | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F16, F32, F32, F32, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, D0ElementOp, 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>;
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,
CDataType,
EDataType,
GemmAccDataType,
AElementOp,
BElementOp,
......@@ -88,9 +95,9 @@ using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
using DeviceNormalizeInstance =
ck::tensor_operation::device::Device5AryElementwise<CDataType,
ReduceDataType,
ReduceDataType,
ck::tensor_operation::device::Device5AryElementwise<EDataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType,
......@@ -124,41 +131,31 @@ auto f_host_tensor_descriptor2d =
}
};
template <typename CDataType,
typename ReduceDataType,
typename AccDataType,
typename BiasDataType,
typename D0DataType,
typename A_functor,
typename B_functor,
typename C_functor,
typename C1_functor>
void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
const Tensor<ADataType>& a_m_k,
const Tensor<ADataType>& b_k_n,
const Tensor<BiasDataType>& bias_n,
const Tensor<D0DataType>& c1_m_n,
const Tensor<BDataType>& b_k_n,
const Tensor<D0DataType>& bias_n,
const Tensor<D1DataType>& d1_m_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<GammaDataType>& beta_n,
A_functor a_element_op,
B_functor b_element_op,
C_functor c_element_op,
C1_functor c1_element_op,
const Tensor<BetaDataType>& beta_n,
AElementOp a_element_op,
BElementOp b_element_op,
CDEElementOp cde_element_op,
int M,
int N)
{
int StrideC = N;
Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<ReduceDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
auto averageOpInst = UnaryDivElementOp{N};
int StrideE = N;
Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<R1DataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
auto averageOpInst = Div{N};
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_gemm.MakeArgument(a_m_k, b_k_n, e_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
......@@ -166,38 +163,32 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
for(int m = 0; m < M; ++m)
for(int n = 0; n < N; ++n)
{
AccDataType acc = ck::type_convert<AccDataType>(c_m_n(m, n)) +
ck::type_convert<AccDataType>(bias_n(n));
AccDataType c1 = ck::type_convert<AccDataType>(c1_m_n(m, n));
c_element_op(acc, acc);
c1_element_op(c1, c1);
acc += c1;
c_m_n(m, n) = ck::type_convert<CDataType>(acc);
auto acc = ck::type_convert<GemmAccDataType>(e_m_n(m, n));
cde_element_op(e_m_n(m, n), acc, bias_n(n), d1_m_n(m, n));
}
// reduce_mean and reduce_square_mean
auto reduceSumOpInst = ReduceSumOp{};
auto r0Op = R0ThreadReduceOp{};
auto r1Op = R1ThreadReduceOp{};
for(int m = 0; m < M; ++m)
{
auto mean_acc = reduceSumOpInst.GetIdentityValue<AccDataType>();
auto square_mean_acc = reduceSumOpInst.GetIdentityValue<AccDataType>();
auto mean_acc = r0Op.GetIdentityValue<ReduceAccDataType>();
auto mean_square_acc = r1Op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
AccDataType c_val = ck::type_convert<AccDataType>(c_m_n(m, n));
AccDataType square_c_val = 0;
UnarySquareElementOp{}(square_c_val, c_val);
auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n(m, n));
ReduceAccDataType square_e_val = 0;
Square{}(square_e_val, e_val);
reduceSumOpInst(mean_acc, c_val);
reduceSumOpInst(square_mean_acc, square_c_val);
r0Op(mean_acc, e_val);
r1Op(mean_square_acc, square_e_val);
}
averageOpInst(mean_acc, mean_acc);
averageOpInst(square_mean_acc, square_mean_acc);
mean_m(m) = ck::type_convert<ReduceDataType>(mean_acc);
meanSquare_m(m) = ck::type_convert<ReduceDataType>(square_mean_acc);
averageOpInst(mean_square_acc, mean_square_acc);
mean_m(m) = ck::type_convert<R0DataType>(mean_acc);
meanSquare_m(m) = ck::type_convert<R1DataType>(mean_square_acc);
}
// LayerNorm
......@@ -206,24 +197,25 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
{
for(int n = 0; n < N; ++n)
{
AccDataType out_acc = 0;
NormalizeComputeDataType out_acc = 0;
layerNormInst(out_acc,
ck::type_convert<AccDataType>(c_m_n(m, n)),
ck::type_convert<AccDataType>(mean_m(m)),
ck::type_convert<AccDataType>(meanSquare_m(m)),
ck::type_convert<AccDataType>(gamma_n(n)),
ck::type_convert<AccDataType>(beta_n(n)));
out_m_n(m, n) = ck::type_convert<ReduceDataType>(out_acc);
ck::type_convert<NormalizeComputeDataType>(e_m_n(m, n)),
ck::type_convert<NormalizeComputeDataType>(mean_m(m)),
ck::type_convert<NormalizeComputeDataType>(meanSquare_m(m)),
ck::type_convert<NormalizeComputeDataType>(gamma_n(n)),
ck::type_convert<NormalizeComputeDataType>(beta_n(n)));
out_m_n(m, n) = ck::type_convert<LayerNormOutDataType>(out_acc);
}
}
}
template <typename ADataType,
typename BDataType,
typename CDataType,
typename BiasDataType,
typename EDataType,
typename D0DataType,
typename ReduceDataType,
typename D1DataType,
typename R0DataType,
typename R1DataType,
typename GammaDataType,
typename BetaDataType,
typename NormalizeDataType>
......@@ -231,12 +223,12 @@ void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M,
{
std::size_t gemm_flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(BiasDataType) * M * N +
sizeof(D0DataType) * M * N + sizeof(ReduceDataType) * M +
sizeof(ReduceDataType) * M;
sizeof(EDataType) * M * N + sizeof(D0DataType) * M * N +
sizeof(D0DataType) * M * N + sizeof(R0DataType) * M +
sizeof(R1DataType) * M;
std::size_t normalize_num_byte = sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M +
sizeof(ReduceDataType) * M + sizeof(GammaDataType) * N +
std::size_t normalize_num_byte = sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
sizeof(R1DataType) * M + sizeof(GammaDataType) * N +
sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
......@@ -259,37 +251,37 @@ int main()
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideC = 1024;
ck::index_t StrideD0 = 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<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<D0DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> reduceMean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<ReduceDataType> reduceMeanSquare_m(f_host_tensor_descriptor1d(M, 1));
Tensor<D0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, ELayout{}));
Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> r0_Mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<R1DataType> r1_MeanSquare_m(f_host_tensor_descriptor1d(M, 1));
Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
Tensor<LayerNormOutDataType> layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
bias_n.GenerateTensorValue(GeneratorTensor_3<BiasDataType>{-1, 1});
c1_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-5, 5});
bias_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-5, 5});
gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * c1_m_n.mDesc.GetElementSpaceSize());
DeviceMem reduceMean_device_buf(sizeof(ReduceDataType) *
reduceMean_m.mDesc.GetElementSpaceSize());
DeviceMem reduceMeanSquare_device_buf(sizeof(ReduceDataType) *
reduceMeanSquare_m.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(D0DataType) * bias_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_Mean_device_buf(sizeof(R0DataType) * r0_Mean_m.mDesc.GetElementSpaceSize());
DeviceMem r1_MeanSquare_device_buf(sizeof(R1DataType) *
r1_MeanSquare_m.mDesc.GetElementSpaceSize());
DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
......@@ -298,60 +290,51 @@ int main()
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
bias_device_buf.ToDevice(bias_n.mData.data());
d0_device_buf.ToDevice(c1_m_n.mData.data());
d1_device_buf.ToDevice(d1_m_n.mData.data());
gamma_device_buf.ToDevice(gamma_n.mData.data());
beta_device_buf.ToDevice(beta_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto d_element_op = D0ElementOp{};
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 = {reduceMean_device_buf.GetDeviceBuffer(),
reduceMeanSquare_device_buf.GetDeviceBuffer()};
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, reduce_mean, reduce_mean_square
auto gemmReduce = DeviceGemmBiasAddReduceInstance{};
// Prepare GEMM, mean, mean_square
auto gemmReduce = DeviceOpInstance{};
auto gemmReduce_invoker = gemmReduce.MakeInvoker();
auto gemmReduce_argument = gemmReduce.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
bias_device_buf.GetDeviceBuffer(),
{d0_device_buf.GetDeviceBuffer()},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
{StrideD0},
gemm_element_ops,
{&d_element_op},
reduce_in_element_ops,
reduce_out_element_ops);
auto gemmReduce_argument = gemmReduce.MakeArgument(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{bias_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
{r0_Mean_device_buf.GetDeviceBuffer(), r1_MeanSquare_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(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
reduceMean_device_buf.SetZero();
reduceMeanSquare_device_buf.SetZero();
// init reducetion buffer to 0
r0_Mean_device_buf.SetZero();
r1_MeanSquare_device_buf.SetZero();
// Prepare LayerNorm
std::array<const void*, 5> input = {c_device_buf.GetDeviceBuffer(),
reduceMean_device_buf.GetDeviceBuffer(),
reduceMeanSquare_device_buf.GetDeviceBuffer(),
std::array<const void*, 5> input = {e_device_buf.GetDeviceBuffer(),
r0_Mean_device_buf.GetDeviceBuffer(),
r1_MeanSquare_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {layerNorm_device_buf.GetDeviceBuffer()};
......@@ -361,12 +344,12 @@ int main()
auto normalize_argument = normalize.MakeArgument(input,
output,
{M, N},
{StrideC, 1},
{StrideE, 1},
{1, 0},
{1, 0},
{0, 1},
{0, 1},
{StrideC, 1},
{StrideE, 1},
NormalizeFunctor{});
if(!normalize.IsSupportedArgument(normalize_argument))
......@@ -383,21 +366,20 @@ int main()
{
// verification
Tensor<LayerNormOutDataType> host_layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
host_gemm_layernorm<CDataType, ReduceDataType, ReduceAccDataType>(host_layerNorm_m_n,
a_m_k,
b_k_n,
bias_n,
c1_m_n,
gamma_n,
beta_n,
a_element_op,
b_element_op,
c_element_op,
d_element_op,
M,
N);
f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
host_gemm_layernorm(host_layerNorm_m_n,
a_m_k,
b_k_n,
bias_n,
d1_m_n,
gamma_n,
beta_n,
a_element_op,
b_element_op,
cde_element_op,
M,
N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData,
......@@ -419,10 +401,11 @@ int main()
if(time_kernel)
DumpGemmLayerNormPerf<ADataType,
BDataType,
CDataType,
BiasDataType,
EDataType,
D0DataType,
ReduceDataType,
D1DataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
......
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
View file @ 6c3c06bf
......@@ -9,7 +9,7 @@
#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/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/device_5ary_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......@@ -28,65 +28,73 @@ 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 CDataType = F16;
using GemmAccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ReduceAccDataType = F32;
using ReduceDataType = F32;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*, ReduceDataType*>;
using R0DataType = F32;
using R1DataType = F32;
using RsDataType = ck::Tuple<R0DataType, R1DataType>;
using GammaDataType = F16;
using BetaDataType = F16;
using LayerNormOutDataType = F16;
using NormalizeComputeDataType = F32;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using 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;
using ReduceOps = ck::Tuple<ReduceSumOp, ReduceSumOp>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using 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;
// Layout
using ALayout = Row;
using BLayout = Col;
using D1Layout = Row;
using ELayout = D1Layout;
// 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 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, 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>;
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,
CDataType,
EDataType,
GemmAccDataType,
AElementOp,
BElementOp,
CElementOp>;
PassThrough>;
using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
using DeviceNormalizeInstance =
ck::tensor_operation::device::Device5AryElementwise<CDataType,
ReduceDataType,
ReduceDataType,
ck::tensor_operation::device::Device5AryElementwise<EDataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType,
......@@ -120,60 +128,54 @@ auto f_host_tensor_descriptor2d =
}
};
template <typename CDataType,
typename ReduceDataType,
typename A_functor,
typename B_functor,
typename C_functor>
void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
const Tensor<ADataType>& a_m_k,
const Tensor<ADataType>& b_k_n,
const Tensor<BDataType>& b_k_n,
const Tensor<GammaDataType>& gamma_n,
const Tensor<BetaDataType>& beta_n,
A_functor a_element_op,
B_functor b_element_op,
C_functor c_element_op,
AElementOp a_element_op,
BElementOp b_element_op,
CDEElementOp c_element_op,
int M,
int N)
{
using out_type = ck::remove_reference_t<decltype(out_m_n(0, 0))>;
int StrideC = N;
Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<ReduceDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
auto averageOpInst = UnaryDivElementOp{N};
int StrideE = N;
Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<R1DataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
auto averageOpInst = Div{N};
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, c_element_op);
ref_gemm.MakeArgument(a_m_k, b_k_n, e_m_n, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
// reduce_mean and reduce_square_mean
auto reduceSumOpInst = ReduceSumOp{};
auto r0Op = R0ThreadReduceOp{};
auto r1Op = R1ThreadReduceOp{};
for(int m = 0; m < M; ++m)
{
auto mean_acc = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
auto square_mean_acc = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
auto mean_acc = r0Op.GetIdentityValue<ReduceAccDataType>();
auto mean_square_acc = r1Op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n(m, n));
auto square_c_val = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n(m, n));
ReduceAccDataType square_e_val = 0;
Square{}(square_e_val, e_val);
UnarySquareElementOp{}(square_c_val, c_val);
reduceSumOpInst(mean_acc, c_val);
reduceSumOpInst(square_mean_acc, square_c_val);
r0Op(mean_acc, e_val);
r1Op(mean_square_acc, square_e_val);
}
averageOpInst(mean_acc, mean_acc);
averageOpInst(square_mean_acc, square_mean_acc);
mean_m(m) = ck::type_convert<ReduceDataType>(mean_acc);
meanSquare_m(m) = ck::type_convert<ReduceDataType>(square_mean_acc);
averageOpInst(mean_square_acc, mean_square_acc);
mean_m(m) = ck::type_convert<R0DataType>(mean_acc);
meanSquare_m(m) = ck::type_convert<R1DataType>(mean_square_acc);
}
// LayerNorm
......@@ -182,22 +184,23 @@ void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
{
for(int n = 0; n < N; ++n)
{
float out_f32 = 0;
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);
NormalizeComputeDataType out_acc = 0;
layerNormInst(out_acc,
ck::type_convert<NormalizeComputeDataType>(e_m_n(m, n)),
ck::type_convert<NormalizeComputeDataType>(mean_m(m)),
ck::type_convert<NormalizeComputeDataType>(meanSquare_m(m)),
ck::type_convert<NormalizeComputeDataType>(gamma_n(n)),
ck::type_convert<NormalizeComputeDataType>(beta_n(n)));
out_m_n(m, n) = ck::type_convert<LayerNormOutDataType>(out_acc);
}
}
}
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ReduceDataType,
typename EDataType,
typename R0DataType,
typename R1DataType,
typename GammaDataType,
typename BetaDataType,
typename NormalizeDataType>
......@@ -205,11 +208,11 @@ void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M,
{
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;
sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
sizeof(R1DataType) * M;
std::size_t normalize_num_btye = sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M +
sizeof(ReduceDataType) * M + sizeof(GammaDataType) * N +
std::size_t normalize_num_btye = sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
sizeof(R1DataType) * M + sizeof(GammaDataType) * N +
sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
......@@ -232,17 +235,17 @@ int main()
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideC = 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<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> reduceMean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<ReduceDataType> reduceMeanSquare_m(f_host_tensor_descriptor1d(M, 1));
Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> r0_Mean_m(f_host_tensor_descriptor1d(M, 1));
Tensor<R1DataType> r1_MeanSquare_m(f_host_tensor_descriptor1d(M, 1));
Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
Tensor<LayerNormOutDataType> layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
......@@ -251,11 +254,10 @@ int main()
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.mDesc.GetElementSpaceSize());
DeviceMem reduceMean_device_buf(sizeof(ReduceDataType) *
reduceMean_m.mDesc.GetElementSpaceSize());
DeviceMem reduceMeanSquare_device_buf(sizeof(ReduceDataType) *
reduceMeanSquare_m.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
DeviceMem r0_Mean_device_buf(sizeof(R0DataType) * r0_Mean_m.mDesc.GetElementSpaceSize());
DeviceMem r1_MeanSquare_device_buf(sizeof(R1DataType) *
r1_MeanSquare_m.mDesc.GetElementSpaceSize());
DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
......@@ -266,40 +268,33 @@ int main()
gamma_device_buf.ToDevice(gamma_n.mData.data());
beta_device_buf.ToDevice(beta_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
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 = {reduceMean_device_buf.GetDeviceBuffer(),
reduceMeanSquare_device_buf.GetDeviceBuffer()};
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, reduce_mean, reduce_mean_square
auto gemmReduce = DeviceGemmReduceInstance{};
// Prepare GEMM, mean, mean_square
auto gemmReduce = DeviceOpInstance{};
auto gemmReduce_invoker = gemmReduce.MakeInvoker();
auto gemmReduce_argument = gemmReduce.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);
auto gemmReduce_argument = gemmReduce.MakeArgument(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
{r0_Mean_device_buf.GetDeviceBuffer(), r1_MeanSquare_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(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
{
......@@ -308,13 +303,13 @@ int main()
"not support this GEMM problem");
}
reduceMean_device_buf.SetZero();
reduceMeanSquare_device_buf.SetZero();
r0_Mean_device_buf.SetZero();
r1_MeanSquare_device_buf.SetZero();
// Prepare LayerNorm
std::array<const void*, 5> input = {c_device_buf.GetDeviceBuffer(),
reduceMean_device_buf.GetDeviceBuffer(),
reduceMeanSquare_device_buf.GetDeviceBuffer(),
std::array<const void*, 5> input = {e_device_buf.GetDeviceBuffer(),
r0_Mean_device_buf.GetDeviceBuffer(),
r1_MeanSquare_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {layerNorm_device_buf.GetDeviceBuffer()};
......@@ -324,12 +319,12 @@ int main()
auto normalize_argument = normalize.MakeArgument(input,
output,
{M, N},
{StrideC, 1},
{StrideE, 1},
{1, 0},
{1, 0},
{0, 1},
{0, 1},
{StrideC, 1},
{StrideE, 1},
NormalizeFunctor{});
if(!normalize.IsSupportedArgument(normalize_argument))
......@@ -346,18 +341,18 @@ int main()
{
// verification
Tensor<LayerNormOutDataType> host_layerNorm_m_n(
f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
host_gemm_layernorm<CDataType, ReduceDataType>(host_layerNorm_m_n,
a_m_k,
b_k_n,
gamma_n,
beta_n,
a_element_op,
b_element_op,
c_element_op,
M,
N);
f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
host_gemm_layernorm(host_layerNorm_m_n,
a_m_k,
b_k_n,
gamma_n,
beta_n,
a_element_op,
b_element_op,
cde_element_op,
M,
N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData,
......@@ -379,8 +374,9 @@ int main()
if(time_kernel)
DumpGemmLayerNormPerf<ADataType,
BDataType,
CDataType,
ReduceDataType,
EDataType,
R0DataType,
R1DataType,
GammaDataType,
BetaDataType,
LayerNormOutDataType>(
......
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/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.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 <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatDsPointer,
typename FloatE,
typename FloatRsPointer,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename QsElementwiseOperation,
typename RsElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename RsGridDescriptor_MBlock_MPerBlock,
typename Block2ETileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_multiple_d_multiple_r_xdl_cshuffle(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatDsPointer p_ds_grid,
FloatE* __restrict__ p_e_grid,
FloatRsPointer p_rs_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const QsElementwiseOperation qs_element_op,
const RsElementwiseOperation rs_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 DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_ds_grid,
p_e_grid,
p_rs_grid,
p_shared,
a_element_op,
b_element_op,
cde_element_op,
qs_element_op,
rs_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock,
rs_grid_desc_mblock_mperblock,
block_2_etile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = p_rs_grid;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = qs_element_op;
ignore = rs_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = rs_grid_desc_mblock_mperblock;
ignore = block_2_etile_map;
#endif
}
} // namespace ck
namespace ck {
namespace tensor_operation {
namespace device {
// GEMM:
// input : A[AK0, M, AK1]
// input : B[AK0, N, AK1]
// input : D0[M, N], D1[M, N], ...
// output : E[M, N]
// output : R0[M], R1[M], ...
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
// R0 = r_op0(Q0), R1 = r_op1(Q1), ...
// Assume:
// D0, D1, ... and E have the same layout
template <typename ALayout,
typename BLayout,
typename DELayout,
typename ADataType,
typename BDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename ReduceAccDataType,
typename RsDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename QsElementwiseOperation,
typename RsElementwiseOperation,
typename ThreadReduceOperations,
typename RsGlobalMemoryDataOperation,
GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1,
index_t BK1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
index_t CDEReduceThreadTransferScalarPerVector_NPerBlock,
index_t RThreadTransferDstScalarPerVector_MPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle
: public DeviceGemmMultipleDMultipleR<ALayout,
BLayout,
DELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
RsDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
QsElementwiseOperation,
RsElementwiseOperation>
{
using DeviceOp = DeviceGemmMultipleDMultipleR_Xdl_CShuffle;
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr index_t NumRTensor = RsDataType::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideA));
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto MPad = M - MRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
}
static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
}();
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto NPad = N - NRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
assert(K % BK1 == 0);
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(NRaw, NPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
assert(KRaw % BK1 == 0);
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
assert(K % BK1 == 0);
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// not pad N or K
assert(KRaw % BK1 == 0);
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
}
static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
{
const auto e_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, DELayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(StrideE, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DELayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(I1, StrideE));
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto MPad = M - MRaw;
const auto NPad = N - NRaw;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(e_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
e_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
e_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return e_grid_desc_mraw_nraw;
}
}
// assume D is packed tensor
static auto MakeRGridDescriptor_M(index_t MRaw)
{
const auto r_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto MPad = M - MRaw;
if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M
return transform_tensor_descriptor(r_grid_desc_mraw,
make_tuple(make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
}
else
{
// not pad M
return r_grid_desc_mraw;
}
}
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(1, 1, 1));
using RGridDesc_M = decltype(MakeRGridDescriptor_M(1));
// GridwiseGemm
using GridwiseGemm = GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ReduceAccDataType,
RsDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
QsElementwiseOperation,
RsElementwiseOperation,
ThreadReduceOperations,
InMemoryDataOperationEnum::Set,
RsGlobalMemoryDataOperation,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
EGridDesc_M_N,
RGridDesc_M,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
CDEReduceThreadTransferScalarPerVector_NPerBlock,
RThreadTransferDstScalarPerVector_MPerBlock,
LoopSched>;
// Argument
struct Argument : public BaseArgument
{
Argument(const void* p_a_grid,
const void* p_b_grid,
std::array<const void*, NumDTensor> p_ds_grid,
void* p_e_grid,
std::array<void*, NumRTensor> p_rs_grid,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
std::array<index_t, NumDTensor> StrideDs,
index_t StrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
QsElementwiseOperation qs_element_op,
RsElementwiseOperation rs_element_op)
: p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
p_ds_grid_{}, // FIXME
p_e_grid_{static_cast<EDataType*>(p_e_grid)},
p_rs_grid_{}, // FIXME
a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N(MRaw, NRaw, StrideE)},
e_grid_desc_mblock_mperblock_nblock_nperblock_{},
r_grid_desc_m_{DeviceOp::MakeRGridDescriptor_M(MRaw)},
rs_grid_desc_mblock_mperblock_{},
block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op},
qs_element_op_{qs_element_op},
rs_element_op_{rs_element_op}
{
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
b_grid_desc_bk0_n_bk1_,
e_grid_desc_m_n_,
r_grid_desc_m_,
block_2_etile_map_))
{
e_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
const auto d_grid_desc_m_n =
DeviceOp::MakeEGridDescriptor_M_N(MRaw, NRaw, StrideDs[i]);
ds_grid_desc_mblock_mperblock_nblock_nperblock_(i) =
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
d_grid_desc_m_n);
});
static_for<0, NumRTensor, 1>{}([&](auto i) {
using RDataType = remove_cvref_t<tuple_element_t<i.value, RsDataType>>;
p_rs_grid_(i) = static_cast<RDataType*>(p_rs_grid[i]);
rs_grid_desc_mblock_mperblock_(i) =
GridwiseGemm::MakeRGridDescriptor_MBlock_MPerBlock(r_grid_desc_m_);
});
}
}
// private:
// pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
typename GridwiseGemm::DsGridPointer p_ds_grid_;
EDataType* p_e_grid_;
typename GridwiseGemm::RsGridPointer p_rs_grid_;
// tensor descriptors
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
StaticallyIndexedArray<
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
NumDTensor>
ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of different
// type from E
EGridDesc_M_N e_grid_desc_m_n_;
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock_;
RGridDesc_M r_grid_desc_m_;
StaticallyIndexedArray<typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock, NumRTensor>
rs_grid_desc_mblock_mperblock_;
// block-to-e-tile map
typename GridwiseGemm::DefaultBlock2ETileMap block_2_etile_map_;
// element-wise op
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
QsElementwiseOperation qs_element_op_;
RsElementwiseOperation rs_element_op_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.e_grid_desc_m_n_,
arg.r_grid_desc_m_,
arg.block_2_etile_map_))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
const index_t grid_size =
arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
auto launch_kernel = [&](auto has_main_k_block_loop) {
constexpr bool has_main_loop = has_main_k_block_loop.value;
const auto kernel = kernel_gemm_multiple_d_multiple_r_xdl_cshuffle<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
typename GridwiseGemm::DsGridPointer,
EDataType,
typename GridwiseGemm::RsGridPointer,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
QsElementwiseOperation,
RsElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
ck::StaticallyIndexedArray<
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
NumDTensor>,
typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
ck::StaticallyIndexedArray<
typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock,
NumRTensor>,
typename GridwiseGemm::DefaultBlock2ETileMap,
has_main_loop>;
return launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_ds_grid_,
arg.p_e_grid_,
arg.p_rs_grid_,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.qs_element_op_,
arg.rs_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.rs_grid_desc_mblock_mperblock_,
arg.block_2_etile_map_);
};
float ave_time = 0;
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
ave_time = launch_kernel(integral_constant<bool, true>{});
}
else
{
ave_time = launch_kernel(integral_constant<bool, false>{});
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
{
return false;
}
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.e_grid_desc_m_n_,
arg.r_grid_desc_m_,
arg.block_2_etile_map_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
std::array<void*, NumRTensor> p_rs,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
std::array<index_t, NumDTensor> StrideDs,
index_t StrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
QsElementwiseOperation qs_element_op,
RsElementwiseOperation rs_element_op)
{
return Argument{p_a,
p_b,
p_ds,
p_e,
p_rs,
MRaw,
NRaw,
KRaw,
StrideA,
StrideB,
StrideDs,
StrideE,
a_element_op,
b_element_op,
cde_element_op,
qs_element_op,
rs_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
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,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
std::array<index_t, NumDTensor> StrideDs,
index_t StrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
QsElementwiseOperation qs_element_op,
RsElementwiseOperation rs_element_op) override
{
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
p_rs,
MRaw,
NRaw,
KRaw,
StrideA,
StrideB,
StrideDs,
StrideE,
a_element_op,
b_element_op,
cde_element_op,
qs_element_op,
rs_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGemmMultipleDMultipleR_Xdl_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock << ", "
<< AK1 << ", "
<< BK1 << ", "
<< getGemmSpecializationString(GemmSpec)
<< ">";
// clang-format on
return str.str();
}
};
} // 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
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp 0 → 100644
View file @ 6c3c06bf
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
template <typename FloatAB,
typename FloatGemmAcc,
typename FloatCShuffle,
typename DsDataType,
typename FloatE,
typename FloatReduceAcc,
typename RsDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename QsElementwiseOperation,
typename RsElementwiseOperation,
typename ThreadReduceOperations,
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
typename RsGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename EGridDesc_M_N,
typename RGridDesc_M,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun,
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun,
index_t BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
index_t CDEReduceThreadTransferScalarPerVector_NPerBlock,
index_t RThreadTransferDstScalarPerVector_MPerBlock,
LoopScheduler LoopSched>
struct GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1
{
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr index_t NumRTensor = RsDataType::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
static constexpr auto AK1 = Number<AK1Value>{};
static constexpr auto BK1 = Number<BK1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(AK0, Number<MPerBlock>{}, AK1),
make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(BK0, Number<NPerBlock>{}, BK1),
make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
}
__host__ __device__ static constexpr auto
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
}
// ck::Tuple<const T0DataType*, const T1DataType*, ...>
template <typename Ts, bool isConst = true>
static constexpr auto MakeTsGridPointer()
{
return generate_tuple(
[&](auto i) {
using T = remove_cvref_t<tuple_element_t<i.value, Ts>>;
if constexpr(isConst)
return static_cast<const T*>(nullptr);
else
return static_cast<T*>(nullptr);
},
Number<Ts::Size()>{});
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for C shuffle in LDS
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
constexpr auto c_block_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
sizeof(FloatAB),
c_block_size * sizeof(FloatCShuffle));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename Block2ETileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const EGridDesc_M_N& e_grid_desc_m_n,
const RGridDesc_M& r_grid_desc_m,
const Block2ETileMap& block_2_etile_map)
{
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
return false;
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
return false;
if(M != r_grid_desc_m.GetLength(I0))
return false;
// check gridwise gemm pipeline
const auto num_k_loop = K / KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
}
if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
{
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
__host__ __device__ static constexpr auto
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
{
const auto M = e_grid_desc_m_n.GetLength(I0);
const auto N = e_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return e_grid_desc_mblock_mperblock_nblock_nperblock;
}
__host__ __device__ static constexpr auto
MakeRGridDescriptor_MBlock_MPerBlock(const RGridDesc_M& r_grid_desc_m)
{
const auto M = r_grid_desc_m.GetLength(I0);
const auto MBlock = M / MPerBlock;
const auto r_grid_desc_mblock_mperblock = transform_tensor_descriptor(
r_grid_desc_m,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1>{}));
return r_grid_desc_mblock_mperblock;
}
// return block_id to E matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto
MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
{
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
e_grid_desc_m_n);
}
using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
// Support 2 dimension in the future. Not only M
using RGridDescriptor_MBlock_MPerBlock =
remove_cvref_t<decltype(MakeRGridDescriptor_MBlock_MPerBlock(RGridDesc_M{}))>;
using DefaultBlock2ETileMap =
remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
using DsGridPointer = decltype(MakeTsGridPointer<DsDataType, true>());
using RsGridPointer = decltype(MakeTsGridPointer<RsDataType, false>());
template <bool HasMainKBlockLoop, typename Block2ETileMap>
__device__ static void
Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
DsGridPointer p_ds_grid,
FloatE* __restrict__ p_e_grid,
RsGridPointer p_rs_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op,
const QsElementwiseOperation& qs_element_op,
const RsElementwiseOperation& rs_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 StaticallyIndexedArray<EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
NumDTensor>&
ds_grid_desc_mblock_mperblock_nblock_nperblock, // FIXME: Ds desc may be of different
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
e_grid_desc_mblock_mperblock_nblock_nperblock,
const StaticallyIndexedArray<RGridDescriptor_MBlock_MPerBlock,
NumRTensor>&
rs_grid_desc_mblock_mperblock, // FIXME: Rs desc may be of different
const Block2ETileMap& block_2_etile_map)
{
// FIXME - Share code with other gemm kernel
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
const auto ds_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ds_grid[i],
ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
},
Number<NumDTensor>{});
auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
auto rs_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_rs_grid(i), rs_grid_desc_mblock_mperblock[i].GetElementSpaceSize());
},
Number<NumRTensor>{});
// divide block work by [M, N]
const auto block_work_idx =
block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_etile_map.ValidCTileIndex(
block_work_idx,
make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[K0PerBlock, MPerBlock] is in LDS
// b_mtx[K0PerBlock, NPerBlock] is in LDS
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
constexpr index_t KPack = math::max(
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
FloatAB,
FloatGemmAcc,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack,
LoopSched>();
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
num_k_block_main_loop);
// shuffle C + Ds + reduction + write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatCShuffle*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
FloatCShuffle,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_der_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
// TODO: this should be implemented as a blockwise reduction
// LDS c_reduce_block_desc_mperblock_nperblock
constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_pass_through_transform(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
make_freeze_transform(I0),
make_pass_through_transform(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
static_assert(CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0) *
CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1) ==
BlockSize,
"wrong!");
static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0) ==
0 &&
(CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1) ==
0,
"wrong!");
constexpr index_t mreduce_per_thread =
(CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0);
constexpr index_t nreduce_per_thread =
(CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1);
constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
Sequence<mreduce_per_thread, nreduce_per_thread>{};
// VGPR cde_reduce_thread_desc_mperblock_nperblock
constexpr auto cde_reduce_thread_desc_mperblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
constexpr auto r_thread_desc_mperblock =
make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
constexpr auto r_thread_desc_mblock_mperblock =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
auto e_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
cde_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
// reduce: threadwise copy from LDS to VGPR
constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
CDRThreadTransferClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
const auto c_reduce_thread_cluster_idx =
c_reduce_thread_cluster_desc.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
const auto c_reduce_thread_data_idx_begin =
c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
// To apply D0, D1, ... and reduction.
// Copy c shuffle from LDS back to VGPR
auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
FloatCShuffle,
FloatReduceAcc,
decltype(c_reduce_block_desc_mperblock_nperblock),
decltype(cde_reduce_thread_desc_mperblock_nperblock),
decltype(c_reduce_thread_lengths_mperblock_nperblock),
Sequence<0, 1>,
1,
CDEReduceThreadTransferScalarPerVector_NPerBlock,
1,
true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
// Copy result of reduction back from VGPR to global
auto reduce_tuple_thread_copy_vgpr_to_global = generate_tuple(
[&](auto I) {
auto p_r_grid = p_rs_grid[I];
auto r_element_op = rs_element_op[I];
auto r_grid_desc_mblock_mperblock = rs_grid_desc_mblock_mperblock[I];
return ThreadwiseTensorSliceTransfer_v1r3<
FloatReduceAcc,
remove_pointer_t<decltype(p_r_grid)>,
decltype(r_thread_desc_mblock_mperblock),
decltype(r_grid_desc_mblock_mperblock),
decltype(r_element_op),
Sequence<1, mreduce_per_thread>,
Sequence<0, 1>,
1,
RThreadTransferDstScalarPerVector_MPerBlock,
RsGlobalMemoryDataOperation::At(I),
1,
false>{r_grid_desc_mblock_mperblock,
make_multi_index(block_work_idx[I0], // mblock
c_reduce_thread_data_idx_begin[I0]), // mperblock
r_element_op};
},
Number<NumRTensor>{});
// D0, D1, ..., Dn
constexpr auto cde_reduce_thread_desc_I1_mperblock_I1_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
// FIXME: Decrease usage of VGPR
// Apply pointwise lambda function from multi-source (Global and LDS) into VGPR
auto ds_thread_buf = generate_tuple(
[&](auto) {
return make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
cde_reduce_thread_desc_I1_mperblock_I1_nperblock.GetElementSpaceSize());
},
Number<NumDTensor>{});
// Copy D0, D1, ..., Dn from global to VGPR
auto ds_thread_copy_global_to_vgpr = generate_tuple(
[&](auto I) {
using DDataType = remove_cvref_t<tuple_element_t<I.value, DsDataType>>;
return ThreadwiseTensorSliceTransfer_v2<
DDataType,
FloatReduceAcc,
decltype(ds_grid_desc_mblock_mperblock_nblock_nperblock[I]),
decltype(cde_reduce_thread_desc_I1_mperblock_I1_nperblock),
Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
Sequence<0, 1, 2, 3>,
3,
CDEReduceThreadTransferScalarPerVector_NPerBlock,
1,
true>(ds_grid_desc_mblock_mperblock_nblock_nperblock[I],
make_multi_index(
I0,
m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
I0,
n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
},
Number<NumDTensor>{});
auto e_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
FloatReduceAcc,
FloatE,
decltype(cde_reduce_thread_desc_I1_mperblock_I1_nperblock),
decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
tensor_operation::element_wise::PassThrough,
Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>, // SliceLengths
Sequence<0, 1, 2, 3>, // DimAccessOrder
3, // DstVectorDim
CDEReduceThreadTransferScalarPerVector_NPerBlock,
InMemoryDataOperationEnum::Set,
1,
true>{
e_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(I0,
m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
I0,
n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]),
tensor_operation::element_wise::PassThrough{}};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_der_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to read from LDS
if constexpr(access_id > 0)
block_sync_lds();
// each thread shuffle data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to write to LDS
block_sync_lds();
// Get shuffle data from LDS to VGPR
c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
c_shuffle_block_buf,
cde_reduce_thread_desc_mperblock_nperblock,
make_tuple(I0, I0),
e_thread_buf);
// Global read D0, D1, ...
static_for<0, NumDTensor, 1>{}([&](auto Id) {
auto& d_thread_copy_global_to_vgpr = ds_thread_copy_global_to_vgpr(Id);
d_thread_copy_global_to_vgpr.Run(
ds_grid_desc_mblock_mperblock_nblock_nperblock[Id],
ds_grid_buf[Id],
cde_reduce_thread_desc_I1_mperblock_I1_nperblock,
make_tuple(I0, I0, I0, I0),
ds_thread_buf(Id));
if constexpr(access_id < num_access - 1)
{
// move on D0, D1, ...
constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
d_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
ds_grid_desc_mblock_mperblock_nblock_nperblock[Id], de_global_step);
}
});
// cde_element_op(e, c, d0, d1, ...);
static_for<0, cde_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
[&](auto i) {
const auto c_ds_src_data_refs = concat_tuple_of_reference(
tie(e_thread_buf[i]),
generate_tie(
[&](auto Id) -> const auto& { return ds_thread_buf[Id][i]; },
Number<NumDTensor>{}));
auto e_dst_data_refs = tie(e_thread_buf(i));
unpack2(cde_element_op, e_dst_data_refs, c_ds_src_data_refs);
});
// Global write E
e_thread_copy_vgpr_to_global.Run(cde_reduce_thread_desc_I1_mperblock_I1_nperblock,
make_tuple(I0, I0, I0, I0),
e_thread_buf,
e_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_buf);
if constexpr(access_id < num_access - 1)
{
// move on E
constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
e_thread_copy_vgpr_to_global.MoveDstSliceWindow(
e_grid_desc_mblock_mperblock_nblock_nperblock, de_global_step);
}
// reduction
static_for<0, NumRTensor, 1>{}([&](auto Ir) {
auto r_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
r_thread_desc_mperblock.GetElementSpaceSize());
auto& reduce_thread_copy_vgpr_to_global =
reduce_tuple_thread_copy_vgpr_to_global(Ir);
using ThreadReduceOperation =
remove_cvref_t<decltype(ThreadReduceOperations{}[Ir])>;
using ThreadwiseReduce =
ThreadwiseReduction<FloatReduceAcc,
decltype(cde_reduce_thread_desc_mperblock_nperblock),
decltype(r_thread_desc_mperblock),
ThreadReduceOperation,
false>;
// threadwise reduction
const auto reduce_identityVal =
ThreadReduceOperation::template GetIdentityValue<FloatReduceAcc>();
static_for<0, mreduce_per_thread, 1>{}(
[&](auto I) { r_thread_buf(I) = reduce_identityVal; });
static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
constexpr auto offset =
Number<cde_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
make_tuple(im, in))>{};
qs_element_op[Ir](e_thread_buf(offset), e_thread_buf(offset));
});
});
ThreadwiseReduce::Reduce(e_thread_buf, r_thread_buf);
// gridwise reduction
reduce_thread_copy_vgpr_to_global.Run(r_thread_desc_mblock_mperblock,
make_tuple(I0, I0),
r_thread_buf,
rs_grid_desc_mblock_mperblock[Ir],
rs_grid_buf(Ir));
if constexpr(access_id < num_access - 1)
{
// move on R0, R1, ...
constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
rs_grid_desc_mblock_mperblock[Ir],
make_tuple(de_global_step[I0], de_global_step[I1]));
}
});
}); // copy c, d, e + reduction
} // shuffle C + Ds + reduction + write out
} // Run
};
} // namespace ck
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