Skip to content

GitLab

Commit 289f15de authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into wmma_gemm

parents 9bd44685 d58b7f51
Hide whitespace changes
Inline Side-by-side
Showing with 1237 additions and 108 deletions
CMakeLists.txt
View file @ 289f15de
......@@ -262,6 +262,16 @@ rocm_package_setup_component(tests
PACKAGE_NAME tests # Prevent -static suffix on package name
)
rocm_package_setup_component(examples
LIBRARY_NAME composablekernel
PACKAGE_NAME examples
)
rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel
PACKAGE_NAME ckProfiler
)
add_subdirectory(library)
add_subdirectory(example)
add_subdirectory(test)
......
Jenkinsfile
View file @ 289f15de
......@@ -618,9 +618,9 @@ pipeline {
stage('Clang Format') {
agent{ label rocmnode("nogpu") }
environment{
execute_cmd = "find .. -iname \'*.h\' \
-o -iname \'*.hpp\' \
-o -iname \'*.cpp\' \
execute_cmd = "find .. -not -path \'*.git*\' -iname \'*.h\' \
-o -not -path \'*.git*\' -iname \'*.hpp\' \
-o -not -path \'*.git*\' -iname \'*.cpp\' \
-o -iname \'*.h.in\' \
-o -iname \'*.hpp.in\' \
-o -iname \'*.cpp.in\' \
......
client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
View file @ 289f15de
add_custom_target(client_gemm_fastgelu_examples)
add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_add_fastgelu gemm_add_fastgelu.cpp)
target_link_libraries(client_gemm_add_fastgelu PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_fastgelu gemm_fastgelu.cpp)
target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_operations)
add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
client_gemm_fastgelu)
client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddFastGelu;
using ADataType = F16;
using BDataType = F16;
using D0DataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using ELayout = Row;
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[])
{
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD0 = 0;
ck::index_t StrideE = 4096;
if(argc == 1)
{
// use default case
}
else if(argc == 8)
{
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]);
StrideD0 = std::stoi(argv[6]);
StrideE = std::stoi(argv[8]);
}
else
{
printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, 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(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
f_matrix_space_size(M, N, StrideD0, D0Layout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddFastGelu>;
// 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{};
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*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
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(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;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = FastGelu;
using ADataType = F16;
using BDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
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[])
{
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideE = 4096;
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[8]);
}
else
{
printf("arg1 to 6: 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(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<>,
ELayout,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::FastGelu>;
// 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{};
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(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
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(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;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/04_contraction/contraction_bilinear.cpp
View file @ 289f15de
......@@ -12,6 +12,7 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
#include "ck/library/utility/numeric.hpp"
using F32 = float;
......@@ -192,20 +193,14 @@ int main(int argc, char* argv[])
{
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 M = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
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 N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
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>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
......
client_example/04_contraction/contraction_scale.cpp
View file @ 289f15de
......@@ -12,6 +12,7 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
#include "ck/library/utility/numeric.hpp"
using F32 = float;
......@@ -178,20 +179,14 @@ int main(int argc, char* argv[])
{
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 M = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
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 N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
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>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
......
client_example/07_conv2d_fwd/CMakeLists.txt deleted 100644 → 0
View file @ 9bd44685
add_executable(client_conv2d_fwd conv2d_fwd.cpp)
target_link_libraries(client_conv2d_fwd PRIVATE composable_kernel::device_operations)
client_example/07_grouped_conv2d_fwd/CMakeLists.txt 0 → 100644
View file @ 289f15de
add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_operations)
client_example/07_grouped_conv2d_fwd/grouped_conv2d_fwd.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 28;
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()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1};
std::partial_sum(rbegin(in_lengths),
std::prev(rend(in_lengths)),
std::next(rbegin(in_strides)),
std::multiplies<>{});
std::partial_sum(rbegin(wei_lengths),
std::prev(rend(wei_lengths)),
std::next(rbegin(wei_strides)),
std::multiplies<>{});
std::partial_sum(rbegin(out_lengths),
std::prev(rend(out_lengths)),
std::next(rbegin(out_strides)),
std::multiplies<>{});
// transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
std::rotate(
rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
std::rotate(
rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
std::rotate(
rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
std::rotate(
rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
std::rotate(
rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
std::rotate(
rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
PassThrough,
PassThrough,
PassThrough>;
// 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;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 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(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
sizeof(WeiDataType) * G * K * Y * X * C +
sizeof(OutDataType) * G * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
client_example/09_quantization/CMakeLists.txt
View file @ 289f15de
add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations)
client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using BiasDataType = int32_t;
using RequantScaleDataType = float;
using OutDataType = int8_t;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_K;
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4;
static constexpr ck::index_t K = 64;
static constexpr ck::index_t C = 32;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Hi = 71;
static constexpr ck::index_t Wi = 71;
static constexpr ck::index_t Ho = 36;
static constexpr ck::index_t Wo = 36;
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[])
{
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout, RequantScaleLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<BiasDataType, RequantScaleDataType>,
OutDataType,
PassThrough,
PassThrough,
OutElementOp>;
// 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;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 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(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths, requant_scale_lengths},
{bias_strides, requant_scale_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr =
op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths, requant_scale_lengths},
{bias_strides, requant_scale_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
\ No newline at end of file
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ 289f15de
......@@ -6,7 +6,7 @@
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_bias_forward_perlayer_quantization.hpp"
#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......
client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using RequantScaleDataType = float;
using OutDataType = int8_t;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4;
static constexpr ck::index_t K = 64;
static constexpr ck::index_t C = 32;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Hi = 71;
static constexpr ck::index_t Wi = 71;
static constexpr ck::index_t Ho = 36;
static constexpr ck::index_t Wo = 36;
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[])
{
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C};
std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<RequantScaleLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<RequantScaleDataType>,
OutDataType,
PassThrough,
PassThrough,
OutElementOp>;
// 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;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 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(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{requant_scale.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{requant_scale_lengths},
{requant_scale_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{},
{},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
\ No newline at end of file
client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ 289f15de
......@@ -6,7 +6,7 @@
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_perlayer_quantization.hpp"
#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......
client_example/09_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp → client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp
View file @ 289f15de
......@@ -98,16 +98,18 @@ int main()
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdData<NumDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
PassThrough,
PassThrough,
PassThrough>;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
PassThrough,
PassThrough,
PassThrough>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
......@@ -126,15 +128,18 @@ int main()
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
......@@ -189,15 +194,18 @@ int main()
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
......
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt 0 → 100644
View file @ 289f15de
add_executable(client_grouped_conv2d_bwd_weight grouped_conv2d_bwd_weight.cpp)
target_link_libraries(client_grouped_conv2d_bwd_weight PRIVATE composable_kernel::device_operations)
client_example/07_conv2d_fwd/conv2d_fwd.cpp → client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight.cpp
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <vector>
#include <iterator>
#include <numeric>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/convolution_forward.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......@@ -15,21 +17,22 @@ using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NHWC;
using WeiLayout = ck::tensor_layout::convolution::KYXC;
using OutLayout = ck::tensor_layout::convolution::NHWK;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 32;
static constexpr ck::index_t C = 3;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Hi = 224;
static constexpr ck::index_t Wi = 224;
static constexpr ck::index_t Ho = 113;
static constexpr ck::index_t Wo = 113;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 28;
struct SimpleDeviceMem
{
......@@ -47,30 +50,33 @@ struct SimpleDeviceMem
void* p_mem_;
};
int main(int argc, char* argv[])
int main()
{
std::vector<ck::index_t> in_spatial_lengths{Hi, Wi};
std::vector<ck::index_t> filter_spatial_lengths{Y, X};
std::vector<ck::index_t> out_spatial_lengths{Ho, Wo};
std::vector<ck::index_t> filter_strides{2, 2};
std::vector<ck::index_t> filter_dilations{1, 1};
std::vector<ck::index_t> input_left_pads{2, 2};
std::vector<ck::index_t> input_right_pads{2, 2};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceConvFwd<NumDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
PassThrough,
PassThrough,
PassThrough>;
std::array<ck::index_t, NumDimSpatial> input_spatial_lengths{Hi, Wi};
std::array<ck::index_t, NumDimSpatial> filter_spatial_lengths{Y, X};
std::array<ck::index_t, NumDimSpatial> output_spatial_lengths{Ho, Wo};
std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1};
std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
ck::index_t split_k = 2;
SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NumDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
PassThrough,
PassThrough,
PassThrough>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
......@@ -92,19 +98,21 @@ int main(int argc, char* argv[])
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
out.GetDeviceBuffer(),
G,
N,
K,
C,
in_spatial_lengths,
input_spatial_lengths,
filter_spatial_lengths,
out_spatial_lengths,
filter_strides,
filter_dilations,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
PassThrough{},
split_k);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
......@@ -112,10 +120,10 @@ int main(int argc, char* argv[])
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * C +
sizeof(WeiDataType) * K * Y * X * C +
sizeof(OutDataType) * N * Ho * Wo * K;
std::size_t flop = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
sizeof(WeiDataType) * G * K * Y * X * C +
sizeof(OutDataType) * G * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
......@@ -134,10 +142,16 @@ int main(int argc, char* argv[])
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
......@@ -149,21 +163,22 @@ int main(int argc, char* argv[])
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
out.GetDeviceBuffer(),
G,
N,
K,
C,
in_spatial_lengths,
input_spatial_lengths,
filter_spatial_lengths,
out_spatial_lengths,
filter_strides,
filter_dilations,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
PassThrough{},
split_k);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
......@@ -172,6 +187,4 @@ int main(int argc, char* argv[])
std::cout << "Done" << std::endl;
}
return 0;
}
\ No newline at end of file
}
client_example/12_elementwise_normalization/CMakeLists.txt 0 → 100644
View file @ 289f15de
add_executable(client_elementwise_layernorm2d elementwise_layernorm2d.cpp)
target_link_libraries(client_elementwise_layernorm2d PRIVATE composable_kernel::device_operations)
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