Skip to content

GitLab

Commit e599063f authored by illsilin's avatar illsilin
Browse files

sync from the public repo

parents 5dbbf5d6 566b6480
Hide whitespace changes
Inline Side-by-side
Showing with 1734 additions and 46 deletions
client_example/30_gemm_bf16Aint8B/gemm_xdl_multiply_bf16_i8.cpp 0 → 100644
View file @ e599063f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#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_gemm_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using DsDataType = ck::Tuple<B1DataType>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<B1Layout>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Multiply;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
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_;
};
// clang-format on
int main(int argc, char* argv[])
{
// GEMM shape
ck::index_t M = 4096;
ck::index_t N = 768;
ck::index_t K = 6144;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideE = N;
if(argc == 1)
{
// use default case
}
else if(argc == 7)
{
M = std::stoi(argv[1]);
N = std::stoi(argv[2]);
K = std::stoi(argv[3]);
StrideA = std::stoi(argv[4]);
StrideB = std::stoi(argv[5]);
StrideE = std::stoi(argv[6]);
}
else
{
printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideE\n");
exit(0);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
f_matrix_space_size(M, K, StrideA, A0Layout{}));
SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
f_matrix_space_size(K, N, StrideB, B0Layout{}));
SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 1;
constexpr ck::index_t NumDTensor = 1;
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
BsLayout,
DsLayout,
Row,
AsDataType,
BsDataType,
DsDataType,
BF16,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
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(
std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, NumATensor>{StrideA},
std::array<ck::index_t, NumBTensor>{StrideB},
std::array<ck::index_t, NumDTensor>{0},
StrideE,
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});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * 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/31_grouped_gemm_bf16Aint8B/CMakeLists.txt 0 → 100644
View file @ e599063f
if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf16") OR NOT DEFINED DTYPES))
add_executable(client_grouped_gemm_bias_fastgelu_bf16_i8_bf16 grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_fastgelu_bf16_i8_bf16 grouped_gemm_fastgelu_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_multiply_bf16_i8_bf16 grouped_gemm_multiply_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_bf16_i8_bf16 grouped_gemm_xdl_bf16_i8.cpp)
target_link_libraries(client_grouped_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
endif()
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp 0 → 100644
View file @ e599063f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#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_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<D0DataType>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = AddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
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_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
d0_tensors_device, c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
d0_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 1;
using GroupedGemmKernelArgument = ck::tensor_operation::device::
GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
d0_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
gemm_descs.push_back(
{sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {0}, 1});
grouped_gemm_kernel_args_.push_back(
{std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
b1_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
std::array<ck::index_t, NumDTensor>{0},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmMultiABDFixedNK<AsLayout,
BsLayout,
DsLayout,
Row,
AsDataType,
BsDataType,
DsDataType,
BF16,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
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];
std::vector<std::array<const void*, NumATensor>> p_As = {};
std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
op_ptr->SetElementwiseOps(
argument_ptr.get(), a_element_op, b_element_op, cde_element_op);
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
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 true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(32 + rand() % 32);
problem_size.Ns.push_back(1024);
problem_size.Ks.push_back(512);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
}
return !run_grouped_gemm(problem_size, config);
}
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_fastgelu_xdl_bf16_i8.cpp 0 → 100644
View file @ e599063f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#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_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = FastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
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_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 0;
using GroupedGemmKernelArgument = ck::tensor_operation::device::
GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
gemm_descs.push_back(
{sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {}, 1});
grouped_gemm_kernel_args_.push_back(
{std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
b1_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumDTensor>{},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
std::array<ck::index_t, NumDTensor>{},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmMultiABDFixedNK<AsLayout,
BsLayout,
DsLayout,
Row,
AsDataType,
BsDataType,
DsDataType,
BF16,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
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];
std::vector<std::array<const void*, NumATensor>> p_As = {};
std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
op_ptr->SetElementwiseOps(
argument_ptr.get(), a_element_op, b_element_op, cde_element_op);
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
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 true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(32 + rand() % 32);
problem_size.Ns.push_back(1024);
problem_size.Ks.push_back(512);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
}
return !run_grouped_gemm(problem_size, config);
}
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp 0 → 100644
View file @ e599063f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#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_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multply.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using B0DataType = I8;
using B1DataType = BF16;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = BF16;
using DsDataType = ck::Tuple<B1DataType, D0DataType>;
using EDataType = BF16;
using A0Layout = Row;
using B0Layout = Row;
using B1Layout = B0Layout;
using D0Layout = Row;
using DsLayout = ck::Tuple<B0Layout, D0Layout>;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MultiplyAddFastGelu;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
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_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
d0_tensors_device, c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
d0_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumDTensor = 2;
using GroupedGemmKernelArgument =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(A0DataType) * problem_size.Ms[i] * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
d0_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
gemm_descs.push_back({problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
problem_size.stride_Cs[i],
{0, 0}});
grouped_gemm_kernel_args_.push_back(
{a0_tensors_device[i]->GetDeviceBuffer(),
b0_tensors_device[i]->GetDeviceBuffer(),
{b1_tensors_device[i]->GetDeviceBuffer(), d0_tensors_device[i]->GetDeviceBuffer()},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
{0, 0},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmTileLoop<A0Layout,
B0Layout,
DsLayout,
ELayout,
A0DataType,
B0DataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
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];
std::vector<const void*> p_As = {};
std::vector<const void*> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, 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()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
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 true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(1 + rand() % 1024);
problem_size.Ns.push_back(6144);
problem_size.Ks.push_back(4096);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
<< problem_size.Ks[i] << std::endl;
}
return !run_grouped_gemm(problem_size, config);
}
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_xdl_bf16_i8.cpp 0 → 100644
View file @ e599063f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#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_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_tile_loop_multply.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using B0DataType = I8;
using B1DataType = BF16;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<B1DataType>;
using EDataType = BF16;
using A0Layout = Row;
using B0Layout = Row;
using B1Layout = B0Layout;
using D0Layout = Row;
using DsLayout = ck::Tuple<B1Layout>;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Multiply;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
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_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumDTensor = 1;
using GroupedGemmKernelArgument =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(A0DataType) * problem_size.Ms[i] * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
gemm_descs.push_back({problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
problem_size.stride_Cs[i],
{0}});
grouped_gemm_kernel_args_.push_back({a0_tensors_device[i]->GetDeviceBuffer(),
b0_tensors_device[i]->GetDeviceBuffer(),
{b1_tensors_device[i]->GetDeviceBuffer()},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
problem_size.stride_As[i],
problem_size.stride_Bs[i],
{0},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmTileLoop<A0Layout,
B0Layout,
DsLayout,
ELayout,
A0DataType,
B0DataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
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];
std::vector<const void*> p_As = {};
std::vector<const void*> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_As, p_Bs, p_Ds, p_Cs, gemm_descs, 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()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
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 true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(1 + rand() % 1024);
problem_size.Ns.push_back(4096);
problem_size.Ks.push_back(4096);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
<< problem_size.Ks[i] << std::endl;
}
return !run_grouped_gemm(problem_size, config);
}
client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_xdl_bf16_i8.cpp 0 → 100644
View file @ e599063f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <iomanip>
#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_grouped_gemm_multi_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
#include "ck/host_utility/hip_check_error.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = BF16;
using AsDataType = ck::Tuple<A0DataType>;
using B0DataType = I8;
using B1DataType = BF16;
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
using AccDataType = F32;
using CShuffleDataType = BF16;
using D0DataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using A0Layout = Row;
using AsLayout = ck::Tuple<A0Layout>;
using B0Layout = Row;
using B1Layout = B0Layout;
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
using D0Layout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using Multiply = ck::tensor_operation::element_wise::Multiply;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = Multiply;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
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_;
};
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int k_batch = 1;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
auto group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
gemm_descs.reserve(group_count);
int sum_of_m = 0;
using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
c_tensors_device;
a0_tensors_device.reserve(group_count);
b0_tensors_device.reserve(group_count);
b1_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < group_count; i++)
{
sum_of_m += problem_size.Ms[i];
}
constexpr ck::index_t NumATensor = 1;
constexpr ck::index_t NumBTensor = 2;
constexpr ck::index_t NumDTensor = 0;
using GroupedGemmKernelArgument = ck::tensor_operation::device::
GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
a0_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
b1_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
c_tensors_device.emplace_back(
std::make_unique<SimpleDeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
gemm_descs.push_back(
{sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {}, 1});
grouped_gemm_kernel_args_.push_back(
{std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
b1_tensors_device[i]->GetDeviceBuffer()},
std::array<const void*, NumDTensor>{},
c_tensors_device[i]->GetDeviceBuffer(),
problem_size.Ms[i],
problem_size.Ns[i],
problem_size.Ks[i],
std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
std::array<ck::index_t, NumDTensor>{},
problem_size.stride_Cs[i]});
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmMultiABDFixedNK<AsLayout,
BsLayout,
DsLayout,
Row,
AsDataType,
BsDataType,
DsDataType,
BF16,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
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];
std::vector<std::array<const void*, NumATensor>> p_As = {};
std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
std::vector<void*> p_Cs = {};
auto argument_ptr = op_ptr->MakeArgumentPointer(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
SimpleDeviceMem gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
op_ptr->SetElementwiseOps(
argument_ptr.get(), a_element_op, b_element_op, cde_element_op);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
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 true;
}
int main(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(1 + rand() % 1024);
problem_size.Ns.push_back(4096);
problem_size.Ks.push_back(4096);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ns[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
<< problem_size.Ks[i] << std::endl;
}
return !run_grouped_gemm(problem_size, config);
}
client_example/CMakeLists.txt
View file @ e599063f
......@@ -48,7 +48,25 @@ else()
endif()
endif()
find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations device_contraction_operations device_reduction_operations)
if (GPU_TARGETS)
if (GPU_TARGETS MATCHES "gfx9")
add_definitions(-DCK_USE_XDL)
set(CK_USE_XDL "ON")
endif()
if (GPU_TARGETS MATCHES "gfx11")
add_definitions(-DCK_USE_WMMA)
set(CK_USE_WMMA "ON")
endif()
else()
add_definitions(-DCK_USE_WMMA -DCK_USE_XDL)
set(CK_USE_XDL "ON")
set(CK_USE_WMMA "ON")
endif()
find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations device_reduction_operations)
if(GPU_TARGETS MATCHES "gfx9")
find_package(composable_kernel COMPONENTS device_contraction_operations)
endif()
find_package(hip REQUIRED PATHS /opt/rocm)
message(STATUS "Build with HIP ${hip_VERSION}")
......
cmake/EnableCompilerWarnings.cmake
View file @ e599063f
......@@ -95,6 +95,7 @@ else()
-Wno-weak-vtables
-Wno-covered-switch-default
-Wno-unsafe-buffer-usage
-Wno-unused-lambda-capture
)
else()
if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "GNU" AND ${COMPILER} MATCHES "CXX")
......
docs/what-is-ck.rst → docs/conceptual/what-is-ck.rst
View file @ e599063f
......@@ -20,7 +20,7 @@ CK utilizes two concepts to achieve performance portability and code maintainabi
* Algorithm complexity reduction for complex ML operators using an innovative technique called
"Tensor Coordinate Transformation".
.. image:: data/ck_component.png
.. image:: ../data/ck_component.png
:alt: CK Components
......@@ -36,6 +36,6 @@ The CK library is structured into 4 layers:
It also includes a simple wrapper component used to perform tensor transform operations more easily and with fewer lines of code.
.. image:: data/ck_layer.png
.. image:: ../data/ck_layer.png
:alt: CK Layers
\ No newline at end of file
docs/index.rst
View file @ e599063f
......@@ -12,28 +12,27 @@ The Composable Kernel (CK) library provides a programming model for writing perf
The CK documentation is structured as follows:
.. card:: Conceptual
.. grid:: 2
:gutter: 3
* :ref:`what-is-ck`
.. grid-item-card:: Installation
.. card:: Installation
* :ref:`docker-hub`
* :ref:`docker-hub`
.. grid-item-card:: Conceptual
.. card:: Tutorial
* :ref:`what-is-ck`
* :ref:`hello-world`
.. grid-item-card:: API reference
.. card:: API reference
* :ref:`supported-primitives`
* :ref:`api-reference`
* :ref:`wrapper`
* :ref:`supported-primitives`
* :ref:`api-reference`
* :ref:`wrapper`
.. grid-item-card:: Tutorial
.. card:: Contributing to CK
* :ref:`hello-world`
* :ref:`contributing-to`
To contribute to the documentation refer to `Contributing to ROCm <https://rocm.docs.amd.com/en/latest/contribute/index.html>`_.
To contribute to the documentation refer to `Contributing to ROCm <https://rocm.docs.amd.com/en/latest/contribute/contributing.html>`_.
You can find licensing information on the `Licensing <https://rocm.docs.amd.com/en/latest/about/license.html>`_ page.
docs/license.md deleted 100644 → 0
View file @ 5dbbf5d6
```{include} ../LICENSE.md
```
docs/license.rst 0 → 100644
View file @ e599063f
.. meta::
:description: Composable Kernel documentation and API reference library
:keywords: composable kernel, CK, ROCm, API, documentation
.. _license:
********************************************************************
License
********************************************************************
.. include:: ../LICENSE
\ No newline at end of file
docs/sphinx/_toc.yml.in
View file @ e599063f
......@@ -2,20 +2,35 @@ defaults:
numbered: False
root: index
subtrees:
- entries:
- file: what-is-ck.rst
- caption: Conceptual
entries:
- file: conceptual/what-is-ck.rst
title: What is Composable Kernel?
- file: dockerhub.rst
- caption: Install
entries:
- file: install/dockerhub.rst
title: Docker Hub
- file: tutorial_hello_world.rst
title: Hello World Tutorial
- file: Supported_Primitives_Guide.rst
- caption: CK API Reference
entries:
- file: reference/Supported_Primitives_Guide.rst
title: Supported Primitives
- file: API_Reference_Guide.rst
- file: reference/API_Reference_Guide.rst
title: API Reference
- file: wrapper.rst
- file: reference/wrapper.rst
title: Wrapper
- caption: Tutorial
entries:
- file: tutorial/tutorial_hello_world.rst
title: Hello World Tutorial
- caption: About
entries:
- file: Contributors_Guide.rst
title: Contributing to CK
- file: license.md
- file: license.rst
title: License
\ No newline at end of file
docs/sphinx/requirements.in
View file @ e599063f
rocm-docs-core==0.35.1
rocm-docs-core==1.1.1
sphinxcontrib-bibtex==2.6.2
docs/sphinx/requirements.txt
View file @ e599063f
#
# This file is autogenerated by pip-compile with Python 3.8
# This file is autogenerated by pip-compile with Python 3.10
# by the following command:
#
# pip-compile requirements.in
......@@ -48,12 +48,6 @@ idna==3.4
# via requests
imagesize==1.4.1
# via sphinx
importlib-metadata==6.8.0
# via
# sphinx
# sphinxcontrib-bibtex
importlib-resources==6.1.1
# via rocm-docs-core
jinja2==3.1.2
# via
# myst-parser
......@@ -96,13 +90,9 @@ pygments==2.15.0
# pydata-sphinx-theme
# sphinx
pyjwt[crypto]==2.6.0
# via
# pygithub
# pyjwt
# via pygithub
pynacl==1.5.0
# via pygithub
pytz==2023.3.post1
# via babel
pyyaml==6.0
# via
# myst-parser
......@@ -113,7 +103,7 @@ requests==2.31.0
# via
# pygithub
# sphinx
rocm-docs-core==0.35.1
rocm-docs-core==1.1.1
# via -r requirements.in
six==1.16.0
# via
......@@ -167,7 +157,3 @@ urllib3==1.26.18
# via requests
wrapt==1.15.0
# via deprecated
zipp==3.17.0
# via
# importlib-metadata
# importlib-resources
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