Skip to content

GitLab

Commit 2b27d5fc authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into rosenrodt/gemm-layernorm

parents f689a155 fa9a0a5c
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2125 additions and 637 deletions
example/25_gemm_bias_c_permute/gemm_bias_c_permute_xdl_fp16.cpp 0 → 100644
View file @ 2b27d5fc
// 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_bias_c_permute_xdl.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/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;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmBiasCPermute_Xdl
//######| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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|
//######| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | 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_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DDataType, EDataType, AElementOp, BElementOp, CDEElementOp, 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<1, 32, 1, 8>, 1>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::index_t M0 = 4;
ck::index_t M1 = 32;
ck::index_t M2 = 128;
ck::index_t N0 = 16;
ck::index_t N1 = 256;
// GEMM shape
ck::index_t M = M0 * M1 * M2;
ck::index_t N = N0 * N1;
ck::index_t K = 128;
ck::index_t stride_A = K;
ck::index_t stride_B = K;
#if 1
// E = [M0, N0, M1, N1, M2]
ck::index_t stride_E_M0 = N0 * M1 * N1 * M2;
ck::index_t stride_E_M1 = N1 * M2;
ck::index_t stride_E_M2 = 1;
ck::index_t stride_E_N0 = M1 * N1 * M2;
ck::index_t stride_E_N1 = M2;
// D = [0, N0, 0, N1, 0]
ck::index_t stride_D_M0 = 0;
ck::index_t stride_D_M1 = 0;
ck::index_t stride_D_M2 = 0;
ck::index_t stride_D_N0 = N1;
ck::index_t stride_D_N1 = 1;
#else
// D = [0, 0, 0, N0, N1]
ck::index_t stride_D_M0 = 0;
ck::index_t stride_D_M1 = 0;
ck::index_t stride_D_M2 = 0;
ck::index_t stride_D_N0 = N1;
ck::index_t stride_D_N1 = 1;
// E = [M0, M1, M2, N0, N1]
ck::index_t stride_E_M0 = M1 * M2 * N0 * N1;
ck::index_t stride_E_M1 = M2 * N0 * N1;
ck::index_t stride_E_M2 = N0 * N1;
ck::index_t stride_E_N0 = N1;
ck::index_t stride_E_N1 = 1;
#endif
const ck::tensor_operation::device::DEGridDesc_M0_M1_M2_N0_N1 d_grid_desc{
M0, M1, M2, N0, N1, stride_D_M0, stride_D_M1, stride_D_M2, stride_D_N0, stride_D_N1};
const ck::tensor_operation::device::DEGridDesc_M0_M1_M2_N0_N1 e_grid_desc{
M0, M1, M2, N0, N1, stride_E_M0, stride_E_M1, stride_E_M2, stride_E_N0, stride_E_N1};
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
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=no, 1=yes)\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}));
}
};
auto f_host_de_tensor_descriptor =
[](ck::tensor_operation::device::DEGridDesc_M0_M1_M2_N0_N1 de_grid_desc) {
std::size_t m0 = de_grid_desc.M0_;
std::size_t m1 = de_grid_desc.M1_;
std::size_t m2 = de_grid_desc.M2_;
std::size_t n0 = de_grid_desc.N0_;
std::size_t n1 = de_grid_desc.N1_;
std::size_t stride_m0 = de_grid_desc.stride_M0_;
std::size_t stride_m1 = de_grid_desc.stride_M1_;
std::size_t stride_m2 = de_grid_desc.stride_M2_;
std::size_t stride_n0 = de_grid_desc.stride_N0_;
std::size_t stride_n1 = de_grid_desc.stride_N1_;
return HostTensorDescriptor(
std::vector<std::size_t>({m0, m1, m2, n0, n1}),
std::vector<std::size_t>({stride_m0, stride_m1, stride_m2, stride_n0, stride_n1}));
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, stride_A, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, stride_B, BLayout{}));
Tensor<DDataType> d_m0_m1_m2_n0_n1(f_host_de_tensor_descriptor(d_grid_desc));
Tensor<EDataType> e_m0_m1_m2_n0_n1_host_result(f_host_de_tensor_descriptor(e_grid_desc));
Tensor<EDataType> e_m0_m1_m2_n0_n1_device_result(f_host_de_tensor_descriptor(e_grid_desc));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d_m0_m1_m2_n0_n1: " << d_m0_m1_m2_n0_n1.mDesc << std::endl;
std::cout << "e_m0_m1_m2_n0_n1: " << e_m0_m1_m2_n0_n1_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});
d_m0_m1_m2_n0_n1.GenerateTensorValue(GeneratorTensor_2<DDataType>{-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});
d_m0_m1_m2_n0_n1.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d_m0_m1_m2_n0_n1_device_buf(sizeof(DDataType) *
d_m0_m1_m2_n0_n1.mDesc.GetElementSpace());
DeviceMem e_m0_m1_m2_n0_n1_device_buf(sizeof(EDataType) *
e_m0_m1_m2_n0_n1_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
d_m0_m1_m2_n0_n1_device_buf.ToDevice(d_m0_m1_m2_n0_n1.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
b_k_n_device_buf.GetDeviceBuffer(),
d_m0_m1_m2_n0_n1_device_buf.GetDeviceBuffer(),
e_m0_m1_m2_n0_n1_device_buf.GetDeviceBuffer(),
M,
N,
K,
stride_A,
stride_B,
d_grid_desc,
e_grid_desc,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(DDataType) * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< device_op.GetTypeString() << std::endl;
if(do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m0 = 0; m0 < M0; ++m0)
for(int m1 = 0; m1 < M1; ++m1)
for(int m2 = 0; m2 < M2; ++m2)
for(int n0 = 0; n0 < N0; ++n0)
for(int n1 = 0; n1 < N1; ++n1)
{
int m = m0 * M1 * M2 + m1 * M2 + m2;
int n = n0 * N1 + n1;
cde_element_op(e_m0_m1_m2_n0_n1_host_result(m0, m1, m2, n0, n1),
ck::type_convert<EDataType>(c_m_n(m, n)),
d_m0_m1_m2_n0_n1(m0, m1, m2, n0, n1));
}
e_m0_m1_m2_n0_n1_device_buf.FromDevice(e_m0_m1_m2_n0_n1_device_result.mData.data());
return ck::utils::check_err(e_m0_m1_m2_n0_n1_device_result.mData,
e_m0_m1_m2_n0_n1_host_result.mData)
? 0
: 1;
}
return 0;
}
example/CMakeLists.txt
View file @ 2b27d5fc
......@@ -42,3 +42,4 @@ add_subdirectory(20_convnd_bwd_weight_xdl)
add_subdirectory(21_gemm_layernorm)
add_subdirectory(22_cgemm)
add_subdirectory(23_softmax)
add_subdirectory(25_gemm_bias_c_permute)
include/ck/tensor_operation/gpu/device/device_5ary_elementwise.hpp
View file @ 2b27d5fc
......@@ -10,7 +10,7 @@
#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/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_5ary_Elementwise_1d.hpp"
#include "ck/device_utility/device_prop.hpp"
#include "ck/device_utility/kernel_launch.hpp"
......@@ -35,7 +35,7 @@ template <typename ADataType,
index_t DScalarPerVector,
index_t EScalarPerVector,
index_t FScalarPerVector>
struct Device5AryElementwise : public BaseOperator
struct Device5AryElementwise : public DeviceElementwise<5, 1, NDim, ElementwiseFunctor>
{
static constexpr auto I0 = Number<0>{};
......@@ -268,12 +268,8 @@ struct Device5AryElementwise : public BaseOperator
return true;
};
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
const CDataType* p_c,
const DDataType* p_d,
const EDataType* p_e,
FDataType* p_f,
static auto MakeArgument(std::array<const void*, 5> p_inputs,
std::array<void*, 1> p_outputs,
std::vector<index_t> lengths,
std::vector<index_t> a_strides,
std::vector<index_t> b_strides,
......@@ -283,12 +279,12 @@ struct Device5AryElementwise : public BaseOperator
std::vector<index_t> f_strides,
ElementwiseFunctor functor)
{
return Argument{p_a,
p_b,
p_c,
p_d,
p_e,
p_f,
return Argument{static_cast<const ADataType*>(p_inputs[0]),
static_cast<const BDataType*>(p_inputs[1]),
static_cast<const CDataType*>(p_inputs[2]),
static_cast<const DDataType*>(p_inputs[3]),
static_cast<const EDataType*>(p_inputs[4]),
static_cast<FDataType*>(p_outputs[0]),
lengths,
a_strides,
b_strides,
......@@ -299,40 +295,58 @@ struct Device5AryElementwise : public BaseOperator
functor};
}
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
const void* p_c,
const void* p_d,
const void* p_e,
void* p_f,
std::vector<index_t> lengths,
std::vector<index_t> a_strides,
std::vector<index_t> b_strides,
std::vector<index_t> c_strides,
std::vector<index_t> d_strides,
std::vector<index_t> e_strides,
std::vector<index_t> f_strides,
ElementwiseFunctor functor)
std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, 5> p_inputs,
std::array<void*, 1> p_outputs,
std::vector<index_t> lengths,
std::vector<std::vector<index_t>> input_strides,
std::vector<std::vector<index_t>> output_strides,
ElementwiseFunctor functor) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<const CDataType*>(p_c),
static_cast<const DDataType*>(p_d),
static_cast<const EDataType*>(p_e),
static_cast<FDataType*>(p_f),
return std::make_unique<Argument>(static_cast<const ADataType*>(p_inputs[0]),
static_cast<const BDataType*>(p_inputs[1]),
static_cast<const CDataType*>(p_inputs[2]),
static_cast<const DDataType*>(p_inputs[3]),
static_cast<const EDataType*>(p_inputs[4]),
static_cast<FDataType*>(p_outputs[0]),
lengths,
a_strides,
b_strides,
c_strides,
d_strides,
e_strides,
f_strides,
input_strides[0],
input_strides[1],
input_strides[2],
input_strides[3],
input_strides[4],
output_strides[0],
functor);
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseInvoker> MakeInvokerPointer() { return std::make_unique<Invoker>(); }
};
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>();
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "Device5aryElementwise"
<< "<"
<< "NDim = " << NDim
<< "MPerThread = " << MPerThread
<< "AScalarPerVector = " << AScalarPerVector
<< "BScalarPerVector = " << BScalarPerVector
<< "CScalarPerVector = " << CScalarPerVector
<< "DScalarPerVector = " << DScalarPerVector
<< "EScalarPerVector = " << EScalarPerVector
<< "FScalarPerVector = " << FScalarPerVector
<< ">";
// clang-format on
return str.str();
}
}; // namespace device
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp 0 → 100644
View file @ 2b27d5fc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct DeviceBatchedGemm : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
ck::index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t Batch) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceBatchedGemmPtr = std::unique_ptr<
DeviceBatchedGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_batched_gemm_xdl.hpp
View file @ 2b27d5fc
......@@ -10,7 +10,7 @@
#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.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
#include "ck/device_utility/device_prop.hpp"
......@@ -152,7 +152,7 @@ template <typename ADataType,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector>
struct DeviceBatchedGemmXdl
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
: public DeviceBatchedGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -339,11 +339,11 @@ struct DeviceBatchedGemmXdl
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t BatchCount)
index_t Batch)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
BatchCount_(BatchCount),
Batch_(Batch),
a_grid_desc_k0_m_k1_{
DeviceBatchedGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA)},
b_grid_desc_k0_n_k1_{
......@@ -376,7 +376,7 @@ struct DeviceBatchedGemmXdl
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
index_t BatchCount_;
index_t Batch_;
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
......@@ -420,7 +420,7 @@ struct DeviceBatchedGemmXdl
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Batch_;
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
......@@ -451,7 +451,7 @@ struct DeviceBatchedGemmXdl
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.BatchCount_,
arg.Batch_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
......@@ -485,7 +485,7 @@ struct DeviceBatchedGemmXdl
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.BatchCount_,
arg.Batch_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
......@@ -539,7 +539,7 @@ struct DeviceBatchedGemmXdl
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t BatchCount)
index_t Batch)
{
return Argument{p_a,
p_b,
......@@ -555,7 +555,7 @@ struct DeviceBatchedGemmXdl
a_element_op,
b_element_op,
c_element_op,
BatchCount};
Batch};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -573,7 +573,7 @@ struct DeviceBatchedGemmXdl
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t BatchCount) override
index_t Batch) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......@@ -589,7 +589,7 @@ struct DeviceBatchedGemmXdl
a_element_op,
b_element_op,
c_element_op,
BatchCount);
Batch);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/device_binary_elementwise.hpp
View file @ 2b27d5fc
......@@ -9,6 +9,7 @@
#include "ck/device_utility/device_prop.hpp"
#include "ck/device_utility/kernel_launch.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_binary_elementwise_1d.hpp"
namespace ck {
......@@ -25,7 +26,7 @@ template <typename ADataType,
index_t AScalarPerVector,
index_t BScalarPerVector,
index_t CScalarPerVector>
struct DeviceBinaryElementwise : public BaseOperator
struct DeviceBinaryElementwise : public DeviceElementwise<2, 1, NDim, ElementwiseFunctor>
{
static constexpr auto I0 = Number<0>{};
......@@ -198,27 +199,30 @@ struct DeviceBinaryElementwise : public BaseOperator
return true;
};
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
std::vector<index_t> lengths,
std::vector<index_t> a_strides,
std::vector<index_t> b_strides,
std::vector<index_t> c_strides,
ElementwiseFunctor functor)
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, 2> p_inputs,
std::array<void*, 1> p_outputs,
std::vector<index_t> lengths,
std::vector<std::vector<index_t>> input_strides,
std::vector<std::vector<index_t>> output_strides,
ElementwiseFunctor functor) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
return std::make_unique<Argument>(static_cast<const ADataType*>(p_inputs[0]),
static_cast<const BDataType*>(p_inputs[1]),
static_cast<CDataType*>(p_outputs[0]),
lengths,
a_strides,
b_strides,
c_strides,
input_strides[0],
input_strides[1],
output_strides[0],
functor);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() { return std::make_unique<Invoker>(); }
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>();
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
......@@ -226,7 +230,11 @@ struct DeviceBinaryElementwise : public BaseOperator
// clang-format off
str << "DeviceBinaryElementwise"
<< "<"
<< "NDim = " << NDim
<< "MPerThread = " << MPerThread
<< "AScalarPerVector = " << AScalarPerVector
<< "BScalarPerVector = " << BScalarPerVector
<< "CScalarPerVector = " << CScalarPerVector
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/device_elementwise.hpp 0 → 100644
View file @ 2b27d5fc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <ck::index_t NumInputTensor,
ck::index_t NumOutputTensor,
index_t NDim,
typename ElementwiseFunctor>
struct DeviceElementwise : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, NumInputTensor> p_inputs,
std::array<void*, NumOutputTensor> p_outputs,
std::vector<index_t> lengths,
std::vector<std::vector<index_t>> input_strides,
std::vector<std::vector<index_t>> output_strides,
ElementwiseFunctor functor) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <ck::index_t NumInputTensor,
ck::index_t NumOutputTensor,
index_t NDim,
typename ElementwiseFunctor>
using DeviceElementwisePtr =
std::unique_ptr<DeviceElementwise<NumInputTensor, NumOutputTensor, NDim, ElementwiseFunctor>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_bias_c_permute.hpp 0 → 100644
View file @ 2b27d5fc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
struct DEGridDesc_M0_M1_M2_N0_N1
{
ck::index_t M0_, M1_, M2_, N0_, N1_;
ck::index_t stride_M0_, stride_M1_, stride_M2_, stride_N0_, stride_N1_;
};
// input : A[M, K], B[K, N],
// input : D[M, N], ...
// output : E[M, N]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D)
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceGemmBiasCPermute : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
const void* p_d,
void* p_e,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
DEGridDesc_M0_M1_M2_N0_N1 d_gride_desc,
DEGridDesc_M0_M1_M2_N0_N1 e_gride_desc,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGemmBiasCPermutePtr = std::unique_ptr<
DeviceGemmBiasCPermute<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
View file @ 2b27d5fc
......@@ -9,91 +9,34 @@ namespace ck {
namespace tensor_operation {
namespace device {
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsReduceAccElementwiseOperation>
template <ck::index_t NumDTensor, ck::index_t NumReduce>
struct DeviceGemmReduce : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
const void* p_bias,
std::array<const void*, NumDTensor> p_ds,
void* p_c,
void* p_dxs,
std::array<void*, NumReduce> p_reduces,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
ck::index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsReduceAccElementwiseOperation dxs_out_element_op,
std::array<ck::index_t, NumDTensor> StrideDs,
std::array<void*, 3> gemm_element_ops,
std::array<void*, NumDTensor> d_element_ops,
std::array<void*, NumReduce> reduce_in_element_ops,
std::array<void*, NumReduce> reduce_out_element_ops,
ck::index_t BatchCount = 1) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsReduceAccElementwiseOperation>
using DeviceGemmReducePtr = std::unique_ptr<DeviceGemmReduce<AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsReduceAccElementwiseOperation>>;
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename C1ElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsReduceAccElementwiseOperation>
struct DeviceGemmBiasAddReduce : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
const void* p_c0,
const void* p_c1,
void* p_dxs,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
ck::index_t StrideC,
ck::index_t StrideC1,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
C1ElementwiseOperation c1_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsReduceAccElementwiseOperation dxs_out_element_op,
ck::index_t BatchCount = 1) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename C1ElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsReduceAccElementwiseOperation>
using DeviceGemmBiasAddReducePtr =
std::unique_ptr<DeviceGemmBiasAddReduce<AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
C1ElementwiseOperation,
DxsInElementwiseOperation,
DxsReduceAccElementwiseOperation>>;
template <ck::index_t NumDTensor, ck::index_t NumReduce>
using DeviceGemmReducePtr = std::unique_ptr<DeviceGemmReduce<NumDTensor, NumReduce>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp 0 → 100644
View file @ 2b27d5fc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct DeviceGemmSplitK : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
ck::index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t KBatch) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGemmSplitKPtr = std::unique_ptr<
DeviceGemmSplitK<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp
View file @ 2b27d5fc
......@@ -10,7 +10,7 @@
#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.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp"
#include "ck/device_utility/device_prop.hpp"
......@@ -57,7 +57,7 @@ template <typename ADataType,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector>
struct DeviceGemmXdlSplitK
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
: public DeviceGemmSplitK<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......
include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ 2b27d5fc
......@@ -10,7 +10,7 @@
#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.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp"
#include "ck/device_utility/device_prop.hpp"
......@@ -59,7 +59,7 @@ template <typename ADataType,
typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CBlockTransferScalarPerVector_NWaveNPerXDL>
struct DeviceGemmXdlSplitKCShuffle
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
: public DeviceGemmSplitK<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -420,21 +420,22 @@ struct DeviceGemmXdlSplitKCShuffle
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
sizeof(CDataType)));
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
};
if(has_main_k0_block_loop)
......
include/ck/tensor_operation/gpu/device/device_normalization.hpp 0 → 100644
View file @ 2b27d5fc
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
struct DeviceNormalization : public BaseOperator
{
// inLengths: input tensor extent(s) from high to low dimension
// inStrides: input tensor stride(s) from high to low dimension
// reduceDims: the dimension(s) the normalization operation is applied
// alpha: typeless pointer in host memory storing the alpha scaling value of type AccDataType
// beta: typeless pointer in host memory storing the beta scaling value of type AccDataType
// in_dev: typeless const pointer in device memory storing the input tensor
// out_dev: typeless pointer in device memory storing the output tensor
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<int> reduceDims,
const void* alpha,
const void* beta,
const void* in_dev,
void* out_dev) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
virtual index_t GetRank() const = 0;
virtual index_t GetNumReduceDim() const = 0;
};
using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View file @ 2b27d5fc
......@@ -11,8 +11,8 @@ namespace element_wise {
struct Add
{
template <typename T>
__host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
template <typename Y, typename X0, typename X1>
__host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
template <>
__host__ __device__ constexpr void
......@@ -28,6 +28,13 @@ struct Add
y = x0 + x1;
};
template <>
__host__ __device__ constexpr void
operator()<half_t>(half_t& y, const float& x0, const half_t& x1) const
{
y = type_convert<half_t>(x0) + x1;
};
// Question: should half_t be supported ?
template <>
__host__ __device__ constexpr void
......
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