Skip to content

GitLab

Commit 1deb01b2 authored by Chao Liu's avatar Chao Liu
Browse files

contraction+scale

parent 7b1ad58a
Hide whitespace changes
Inline Side-by-side
Showing with 1140 additions and 4 deletions
client_example/05_contraction_bilinear/CMakeLists.txt → client_example/04_contraction/CMakeLists.txt
View file @ 1deb01b2
add_executable(client_contraction_scale contraction_scale.cpp)
target_link_libraries(client_contraction_scale PRIVATE composable_kernel::device_operations)
add_executable(client_contraction_bilinear contraction_bilinear.cpp) add_executable(client_contraction_bilinear contraction_bilinear.cpp)
target_link_libraries(client_contraction_bilinear PRIVATE composable_kernel::device_operations) target_link_libraries(client_contraction_bilinear PRIVATE composable_kernel::device_operations)
client_example/04_contraction/contraction_scale.cpp 0 → 100644
View file @ 1deb01b2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <numeric>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Scale;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main(int argc, char* argv[])
{
// A[M0, M1, K0, K1]
std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
float scale = 1.f;
if(argc == 1)
{
// use default case
}
else if(argc == 20)
{
const ck::index_t M0 = std::stoi(argv[1]);
const ck::index_t M1 = std::stoi(argv[2]);
const ck::index_t N0 = std::stoi(argv[3]);
const ck::index_t N1 = std::stoi(argv[4]);
const ck::index_t K0 = std::stoi(argv[5]);
const ck::index_t K1 = std::stoi(argv[6]);
a_ms_ks_lengths = {M0, M1, K0, K1};
a_ms_ks_strides = {
std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
b_ns_ks_lengths = {N0, N1, K0, K1};
b_ns_ks_strides = {
std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
e_ms_ns_lengths = {M0, M1, N0, N1};
e_ms_ns_strides = {
std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
scale = std::stof(argv[19]);
}
else
{
printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
printf("arg19: scale\n");
exit(0);
}
auto f_tensor_space_size = [](auto lengths, auto strides) {
std::size_t space_size = 1;
for(std::size_t i = 0; i < lengths.size(); ++i)
{
space_size += (lengths[i] - 1) * strides[i];
}
return space_size;
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) *
f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
SimpleDeviceMem b_device_buf(sizeof(BDataType) *
f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
SimpleDeviceMem e_device_buf(sizeof(EDataType) *
f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{scale};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 0>{},
e_device_buf.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 0>{},
std::array<std::vector<ck::index_t>, 0>{},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
ck::index_t M = std::accumulate(e_ms_ns_lengths.begin(),
e_ms_ns_lengths.begin() + NumDimM,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t N = std::accumulate(e_ms_ns_lengths.begin() + NumDimM,
e_ms_ns_lengths.begin() + NumDimM + NumDimN,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t K = std::accumulate(a_ms_ks_lengths.begin() + NumDimM,
a_ms_ks_lengths.begin() + NumDimM + NumDimK,
ck::index_t{1},
std::multiplies<ck::index_t>{});
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(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: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return 0;
}
client_example/CMakeLists.txt
View file @ 1deb01b2
...@@ -9,4 +9,4 @@ message(STATUS "Build with HIP ${hip_VERSION}") ...@@ -9,4 +9,4 @@ message(STATUS "Build with HIP ${hip_VERSION}")
add_subdirectory(01_gemm) add_subdirectory(01_gemm)
add_subdirectory(02_gemm_add_add_fastgelu) add_subdirectory(02_gemm_add_add_fastgelu)
add_subdirectory(03_gemm_layernorm) add_subdirectory(03_gemm_layernorm)
add_subdirectory(05_contraction_bilinear) add_subdirectory(04_contraction)
example/24_contraction_bilinear/CMakeLists.txt → example/24_contraction/CMakeLists.txt
View file @ 1deb01b2
add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp) add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
example/24_contraction_bilinear/contraction_bilinear_xdl_fp32.cpp → example/24_contraction/contraction_bilinear_xdl_fp32.cpp
View file @ 1deb01b2
...@@ -15,7 +15,6 @@ ...@@ -15,7 +15,6 @@
#include "ck/library/host_tensor/device_memory.hpp" #include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp" #include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp" #include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
......
example/24_contraction/contraction_scale_xdl_fp32.cpp 0 → 100644
View file @ 1deb01b2
// 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/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// clang-format off
using DeviceOpInstanceKKNN = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, F32, F32, F32, F32, DsDataType, F32, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>;
using DeviceOpInstanceKNNN = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, F32, F32, F32, F32, DsDataType, F32, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>;
using DeviceOpInstanceMKNN = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, F32, F32, F32, F32, DsDataType, F32, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>;
using DeviceOpInstanceMNNN = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, F32, F32, F32, F32, DsDataType, F32, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>;
// clang-format on
using DeviceOpInstance = DeviceOpInstanceKKNN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename EDataType,
typename AccDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_ms_ks_{a_ms_ks},
b_ns_ks_{b_ns_ks},
e_ms_ns_{e_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_ms_ks_;
const Tensor<BDataType>& b_ns_ks_;
Tensor<EDataType>& e_ms_ns_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
};
// Invoker
struct Invoker : public ck::tensor_operation::device::BaseInvoker
{
using Argument = ReferenceContraction_M2_N2_K2::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
for(int k1 = 0; k1 < K1; ++k1)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a, static_cast<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
arg.b_element_op_(
v_b, static_cast<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
v_acc += v_a * v_b;
}
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_ms_ns_.mDesc.GetLengths()[0],
arg.e_ms_ns_.mDesc.GetLengths()[1],
arg.e_ms_ns_.mDesc.GetLengths()[2],
arg.e_ms_ns_.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
return 0;
}
float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
{
return true;
}
static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceContraction_M2_N2_K2"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// A[M0, M1, K0, K1]
std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
float scale = 1.f;
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 if(argc == 23)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t M0 = std::stoi(argv[4]);
const ck::index_t M1 = std::stoi(argv[5]);
const ck::index_t N0 = std::stoi(argv[6]);
const ck::index_t N1 = std::stoi(argv[7]);
const ck::index_t K0 = std::stoi(argv[8]);
const ck::index_t K1 = std::stoi(argv[9]);
a_ms_ks_lengths = {M0, M1, K0, K1};
a_ms_ks_strides = {
std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
b_ns_ks_lengths = {N0, N1, K0, K1};
b_ns_ks_strides = {
std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
e_ms_ns_lengths = {M0, M1, N0, N1};
e_ms_ns_strides = {
std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
scale = std::stof(argv[26]);
}
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");
printf("arg4 to 7: M0, M1, N0, N1, K0, K1\n");
printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
printf("arg18 to 21: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
printf("arg22: scale\n");
exit(0);
}
Tensor<ADataType> a_ms_ks(
std::vector<std::size_t>(a_ms_ks_lengths.begin(), a_ms_ks_lengths.end()),
std::vector<std::size_t>(a_ms_ks_strides.begin(), a_ms_ks_strides.end()));
Tensor<BDataType> b_ns_ks(
std::vector<std::size_t>(b_ns_ks_lengths.begin(), b_ns_ks_lengths.end()),
std::vector<std::size_t>(b_ns_ks_strides.begin(), b_ns_ks_strides.end()));
Tensor<EDataType> e_ms_ns_host_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
Tensor<EDataType> e_ms_ns_device_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_ms_ks.mData.data());
b_device_buf.ToDevice(b_ns_ks.mData.data());
// set zero
e_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{scale};
// device operation
auto op = DeviceOpInstance{};
auto invoker = op.MakeInvoker();
auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 0>{},
e_device_buf.GetDeviceBuffer(),
a_ms_ks_lengths,
a_ms_ks_strides,
b_ns_ks_lengths,
b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 0>{},
std::array<std::vector<ck::index_t>, 0>{},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
ck::index_t M = std::accumulate(e_ms_ns_lengths.begin(),
e_ms_ns_lengths.begin() + NumDimM,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t N = std::accumulate(e_ms_ns_lengths.begin() + NumDimM,
e_ms_ns_lengths.begin() + NumDimM + NumDimN,
ck::index_t{1},
std::multiplies<ck::index_t>{});
ck::index_t K = std::accumulate(a_ms_ks_lengths.begin() + NumDimM,
a_ms_ks_lengths.begin() + NumDimM + NumDimK,
ck::index_t{1},
std::multiplies<ck::index_t>{});
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(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, "
<< op.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
c_ms_ns_host_result(m0, m1, n0, n1));
}
}
}
}
return ck::utils::check_err(e_ms_ns_device_result.mData, e_ms_ns_host_result.mData) ? 0 : 1;
}
return 0;
}
example/CMakeLists.txt
View file @ 1deb01b2
...@@ -42,5 +42,5 @@ add_subdirectory(20_convnd_bwd_weight_xdl) ...@@ -42,5 +42,5 @@ add_subdirectory(20_convnd_bwd_weight_xdl)
add_subdirectory(21_gemm_layernorm) add_subdirectory(21_gemm_layernorm)
add_subdirectory(22_cgemm) add_subdirectory(22_cgemm)
add_subdirectory(23_softmax) add_subdirectory(23_softmax)
add_subdirectory(24_contraction_bilinear) add_subdirectory(24_contraction)
add_subdirectory(25_gemm_bias_c_permute) add_subdirectory(25_gemm_bias_c_permute)
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 1deb01b2
...@@ -24,9 +24,25 @@ struct PassThrough ...@@ -24,9 +24,25 @@ struct PassThrough
}; };
}; };
struct Scale
{
__host__ __device__ Scale(float scale) : scale_(scale) {}
template <typename Y, typename X>
__host__ __device__ void operator()(Y& y, const X& x) const;
template <>
__host__ __device__ void operator()<float, float>(float& y, const float& x) const
{
y = scale_ * x;
};
float scale_;
};
struct UnaryDivide struct UnaryDivide
{ {
__host__ __device__ UnaryDivide(const int32_t divider = 1) : divider_(divider){}; __host__ __device__ UnaryDivide(const int32_t divider = 1) : divider_(divider) {}
template <typename T> template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const __host__ __device__ void operator()(T& y, const T& x) const
......
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ 1deb01b2
...@@ -16,6 +16,8 @@ using F32 = float; ...@@ -16,6 +16,8 @@ using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using EMPTY_TUPLE = ck::Tuple<>;
using F16_TUPLE = ck::Tuple<F16>; using F16_TUPLE = ck::Tuple<F16>;
using F16_F16_TUPLE = ck::Tuple<F16, F16>; using F16_F16_TUPLE = ck::Tuple<F16, F16>;
...@@ -25,6 +27,7 @@ using Row = ck::tensor_layout::gemm::RowMajor; ...@@ -25,6 +27,7 @@ using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor; using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
using Bilinear = ck::tensor_operation::element_wise::Bilinear; using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu; using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
......
library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp 0 → 100644
View file @ 1deb01b2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances);
// Contraction + Scale
template <index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename ADataType,
typename BDataType,
typename EDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>>
{
using DeviceOp = DeviceContractionMultipleD<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Scale>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<EDataType, float>)
{
if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
{
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
op_ptrs);
add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 1deb01b2
...@@ -14,6 +14,7 @@ add_subdirectory(gemm_bias_add_reduce) ...@@ -14,6 +14,7 @@ add_subdirectory(gemm_bias_add_reduce)
add_subdirectory(batched_gemm) add_subdirectory(batched_gemm)
add_subdirectory(batched_gemm_reduce) add_subdirectory(batched_gemm_reduce)
add_subdirectory(grouped_gemm) add_subdirectory(grouped_gemm)
add_subdirectory(contraction_scale)
add_subdirectory(contraction_bilinear) add_subdirectory(contraction_bilinear)
add_subdirectory(conv1d_fwd) add_subdirectory(conv1d_fwd)
add_subdirectory(conv2d_fwd) add_subdirectory(conv2d_fwd)
...@@ -36,6 +37,7 @@ add_library(device_operations STATIC ...@@ -36,6 +37,7 @@ add_library(device_operations STATIC
$<TARGET_OBJECTS:device_batched_gemm_instance> $<TARGET_OBJECTS:device_batched_gemm_instance>
$<TARGET_OBJECTS:device_batched_gemm_reduce_instance> $<TARGET_OBJECTS:device_batched_gemm_reduce_instance>
$<TARGET_OBJECTS:device_grouped_gemm_instance> $<TARGET_OBJECTS:device_grouped_gemm_instance>
$<TARGET_OBJECTS:device_contraction_scale_instance>
$<TARGET_OBJECTS:device_contraction_bilinear_instance> $<TARGET_OBJECTS:device_contraction_bilinear_instance>
$<TARGET_OBJECTS:device_conv1d_fwd_instance> $<TARGET_OBJECTS:device_conv1d_fwd_instance>
$<TARGET_OBJECTS:device_conv2d_fwd_instance> $<TARGET_OBJECTS:device_conv2d_fwd_instance>
......
library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt 0 → 100644
View file @ 1deb01b2
# device_contraction_scale_instance
set(DEVICE_CONTRACTION_SCALE_INSTANCE_SOURCE
device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
)
add_library(device_contraction_scale_instance OBJECT ${DEVICE_CONTRACTION_SCALE_INSTANCE_SOURCE})
set_target_properties(device_contraction_scale_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(device_contraction_scale_instance)
library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp 0 → 100644
View file @ 1deb01b2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
// setting Don't use this hack unless absolutely necessary!
// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
#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_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using EMPTY_TUPLE = ck::Tuple<>;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
// k/k/n are the fast changing dimension for A/B/E
using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| 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| Specialization| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances)
{
add_device_operation_instances(
instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp 0 → 100644
View file @ 1deb01b2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
// setting Don't use this hack unless absolutely necessary!
// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
#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_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using EMPTY_TUPLE = ck::Tuple<>;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
// k/n/n are the fast changing dimension for A/B/E
using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| 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| Specialization| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 1, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 1, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 1, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 1, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 1, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 1, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 1, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>
// clang-format on
>;
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances)
{
add_device_operation_instances(
instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp 0 → 100644
View file @ 1deb01b2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
// setting Don't use this hack unless absolutely necessary!
// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
#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_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using EMPTY_TUPLE = ck::Tuple<>;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
// m/k/n are the fast changing dimension for A/B/E
using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| 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| Specialization| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 256, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 256, 16, 1, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 128, 16, 1, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 4, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 4, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 4, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 4, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>
// clang-format on
>;
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances)
{
add_device_operation_instances(
instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp 0 → 100644
View file @ 1deb01b2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
// setting Don't use this hack unless absolutely necessary!
// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
#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_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using EMPTY_TUPLE = ck::Tuple<>;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
// m/n/n are the fast changing dimension for A/B/E
using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance = std::tuple<
// clang-format off
//#####################################| NumDimM| NumDimN| NumDimK| 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| Specialization| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 256, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 256, 16, 1, 1, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 128, 16, 1, 1, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 128, 16, 1, 1, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 64, 16, 1, 1, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 64, 128, 16, 1, 1, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 64, 16, 1, 1, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 64, 128, 16, 1, 1, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 1, 0, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceContractionMultipleD_Xdl_CShuffle< 2, 2, 2, F32, F32, F32, F32, EMPTY_TUPLE, F32, PassThrough, PassThrough, Scale, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>
// clang-format on
>;
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
2,
2,
F32,
F32,
EMPTY_TUPLE,
F32,
PassThrough,
PassThrough,
Scale>>>& instances)
{
add_device_operation_instances(
instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // 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