Skip to content

GitLab

Commit 2732d06c authored by rocking's avatar rocking
Browse files

Merge commit '75891161' into gemm_layernorm_welford

parents dd0255ba 75891161
Hide whitespace changes
Inline Side-by-side
Showing with 1173 additions and 790 deletions
CMakeLists.txt
View file @ 2732d06c
......@@ -29,6 +29,11 @@ if(USE_BITINT_EXTENSION_INT4)
message("CK compiled with USE_BITINT_EXTENSION_INT4 set to ${USE_BITINT_EXTENSION_INT4}")
endif()
## Threads
set(THREADS_PREFER_PTHREAD_FLAG ON)
find_package(Threads REQUIRED)
link_libraries(Threads::Threads)
## C++
enable_language(CXX)
set(CMAKE_CXX_STANDARD 17)
......@@ -78,6 +83,8 @@ if( DEFINED CK_OVERRIDE_HIP_VERSION_PATCH )
message(STATUS "CK_HIP_VERSION_PATCH overriden with ${CK_OVERRIDE_HIP_VERSION_PATCH}")
endif()
message(STATUS "Build with HIP ${HIP_VERSION}")
link_libraries(hip::device)
add_compile_definitions(__HIP_PLATFORM_HCC__=1)
## tidy
include(EnableCompilerWarnings)
......@@ -227,6 +234,7 @@ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include
${PROJECT_SOURCE_DIR}/library/include
${HIP_INCLUDE_DIRS}
)
......
Jenkinsfile
View file @ 2732d06c
......@@ -23,6 +23,22 @@ def getDockerImageName(){
return img
}
def build_compiler(){
def compiler
if (params.BUILD_COMPILER == "hipcc"){
compiler = '/opt/rocm/bin/hipcc'
}
else{
if (params.COMPILER_VERSION == "release"){
compiler = "/opt/rocm/llvm/bin/clang++"
}
else{
compiler = "/llvm-project/build/bin/clang++"
}
}
return compiler
}
def getDockerImage(Map conf=[:]){
env.DOCKER_BUILDKIT=1
def prefixpath = conf.get("prefixpath", "/opt/rocm") // prefix:/opt/rocm
......@@ -103,7 +119,7 @@ def buildDocker(install_prefix){
def cmake_build(Map conf=[:]){
def compiler = conf.get("compiler","/opt/rocm/bin/hipcc")
def compiler = build_compiler()
def config_targets = conf.get("config_targets","check")
def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=undefined -fno-sanitize-recover=undefined " + conf.get("extradebugflags", "")
def build_envs = "CTEST_PARALLEL_LEVEL=4 " + conf.get("build_env","")
......@@ -185,7 +201,6 @@ def buildHipClangJob(Map conf=[:]){
if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 --env GPU_ARCH='${gpu_arch}' "
}
//def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg GPU_ARCH='${gpu_arch}' --build-arg compiler_version='${params.COMPILER_VERSION}' "
def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' "
if (params.COMPILER_VERSION != "release"){
dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
......@@ -467,6 +482,10 @@ pipeline {
name: 'COMPILER_VERSION',
defaultValue: 'ck-9110',
description: 'Specify which version of compiler to use: ck-9110 (default), release, or amd-stg-open.')
string(
name: 'BUILD_COMPILER',
defaultValue: 'hipcc',
description: 'Specify whether to build CK with hipcc (default) or with clang.')
booleanParam(
name: "RUN_FULL_QA",
defaultValue: false,
......@@ -584,8 +603,8 @@ pipeline {
{
agent{ label rocmnode("gfx908")}
environment{
setup_args = """ -D -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DCMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc .. && make -j """
setup_args = """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 -O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
}
steps{
buildHipClangJobAndReboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
......
example/01_gemm/run_gemm_example.inc
View file @ 2732d06c
......@@ -43,30 +43,28 @@ bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
}
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<
#ifdef BUILD_INT4_EXAMPLE
KernelCDataType
#else
CDataType
#endif
>
c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
#ifdef BUILD_INT4_EXAMPLE
DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
c_m_n_device_result.mDesc.GetElementSpaceSize());
const Tensor<KernelADataType> a_m_k_converted(a_m_k);
const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
#else
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
#endif
......@@ -80,13 +78,13 @@ bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(
#ifdef BUILD_INT4_EXAMPLE
reinterpret_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
reinterpret_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
reinterpret_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#else
reinterpret_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
reinterpret_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
reinterpret_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#endif
M,
N,
......@@ -128,13 +126,17 @@ bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
ref_invoker.Run(ref_argument);
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
#ifdef BUILD_INT4_EXAMPLE
const Tensor<CDataType> c_m_n_device_result_converted(c_m_n_device_result);
Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
return ck::utils::check_err(c_m_n_device_result_converted.mData, c_m_n_host_result.mData);
#else
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
#endif
}
......
example/09_convnd_fwd/convnd_fwd_common.hpp
View file @ 2732d06c
......@@ -34,16 +34,16 @@ template <ck::index_t NDimSpatial,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvNDFwdInstance>
int run_grouped_conv_fwd(bool do_verification,
int init_method,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
bool run_grouped_conv_fwd(bool do_verification,
int init_method,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
......@@ -164,10 +164,8 @@ int run_grouped_conv_fwd(bool do_verification,
out_device_buf.FromDevice(out_device.mData.data());
return ck::utils::check_err(
out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f)
? 0
: 1;
out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
}
return 0;
return true;
}
example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
View file @ 2732d06c
......@@ -74,154 +74,6 @@ using DeviceGroupedConvNDFwdInstance =
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
#include "run_convnd_fwd_example.inc"
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1)
{
using InLayout = ctc::GNWC;
using WeiLayout = ctc::GKXC;
using OutLayout = ctc::GNWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<1, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::GNHWC;
using WeiLayout = ctc::GKYXC;
using OutLayout = ctc::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<2, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::GNDHWC;
using WeiLayout = ctc::GKZYXC;
using OutLayout = ctc::GNDHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<3, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
return 0;
}
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
View file @ 2732d06c
......@@ -74,154 +74,6 @@ using DeviceGroupedConvNDFwdInstance =
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
#include "run_convnd_fwd_example.inc"
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1)
{
using InLayout = ctc::GNWC;
using WeiLayout = ctc::GKXC;
using OutLayout = ctc::GNWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<1, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::GNHWC;
using WeiLayout = ctc::GKYXC;
using OutLayout = ctc::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<2, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::GNDHWC;
using WeiLayout = ctc::GKZYXC;
using OutLayout = ctc::GNDHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<3, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
return 0;
}
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
View file @ 2732d06c
......@@ -74,154 +74,6 @@ using DeviceGroupedConvNDFwdInstance =
S<1, 16, 1, 16>,
4>;
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
#include "run_convnd_fwd_example.inc"
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1)
{
using InLayout = ctc::GNWC;
using WeiLayout = ctc::GKXC;
using OutLayout = ctc::GNWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<1, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::GNHWC;
using WeiLayout = ctc::GKYXC;
using OutLayout = ctc::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<2, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::GNDHWC;
using WeiLayout = ctc::GKZYXC;
using OutLayout = ctc::GNDHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<3, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
return 0;
}
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
View file @ 2732d06c
......@@ -74,154 +74,6 @@ using DeviceGroupedConvNDFwdInstance =
S<1, 16, 1, 16>,
1>;
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
#include "run_convnd_fwd_example.inc"
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1)
{
using InLayout = ctc::GNWC;
using WeiLayout = ctc::GKXC;
using OutLayout = ctc::GNWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<1, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::GNHWC;
using WeiLayout = ctc::GKYXC;
using OutLayout = ctc::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<2, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::GNDHWC;
using WeiLayout = ctc::GKZYXC;
using OutLayout = ctc::GNDHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<3, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
return 0;
}
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
View file @ 2732d06c
......@@ -74,154 +74,6 @@ using DeviceGroupedConvNDFwdInstance =
S<1, 64, 1, 4>,
16>;
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
#include "run_convnd_fwd_example.inc"
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1)
{
using InLayout = ctc::GNWC;
using WeiLayout = ctc::GKXC;
using OutLayout = ctc::GNWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<1, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::GNHWC;
using WeiLayout = ctc::GKYXC;
using OutLayout = ctc::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<2, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::GNDHWC;
using WeiLayout = ctc::GKZYXC;
using OutLayout = ctc::GNDHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<3, InLayout, WeiLayout, OutLayout>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
return 0;
}
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
example/09_convnd_fwd/run_convnd_fwd_example.inc 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
bool run_convnd_fwd_example(int argc, char* argv[])
{
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
const auto run = [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto out_layout) {
constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
using InLayout = decltype(in_layout);
using WeiLayout = decltype(wei_layout);
using OutLayout = decltype(out_layout);
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<
ndim_spatial_value,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
};
namespace ctc = ck::tensor_layout::convolution;
if(conv_param.num_dim_spatial_ == 1)
{
return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{});
}
else if(conv_param.num_dim_spatial_ == 2)
{
return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{});
}
else if(conv_param.num_dim_spatial_ == 3)
{
return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
}
return true;
}
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt 0 → 100644
View file @ 2732d06c
add_custom_target(example_convnd_fwd_reduce_xdl)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <cstdlib>
#include <iostream>
#include <iterator>
#include <numeric>
#include <type_traits>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
using BF16 = ck::bhalf_t;
using FP16 = ck::half_t;
using FP32 = float;
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
using I4 = ck::int4_t;
#endif
using I8 = std::int8_t;
using I32 = std::int32_t;
template <typename ALay, typename BLay, typename DELay, typename RLay>
struct LayoutSetting
{
using ALayout = ALay;
using BLayout = BLay;
using DELayout = DELay;
using RLayout = RLay;
};
template <ck::index_t NDimSpatial>
struct LayoutSettingSelector;
namespace ctl = ck::tensor_layout::convolution;
template <>
struct LayoutSettingSelector<1> final : LayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK, ctl::GNW>
{
};
template <>
struct LayoutSettingSelector<2> final : LayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK, ctl::GNHW>
{
};
template <>
struct LayoutSettingSelector<3> final
: LayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK, ctl::GNDHW>
{
};
template <ck::index_t NDimSpatial>
using ALayout = typename LayoutSettingSelector<NDimSpatial>::ALayout;
template <ck::index_t NDimSpatial>
using BLayout = typename LayoutSettingSelector<NDimSpatial>::BLayout;
template <ck::index_t NDimSpatial>
using DELayout = typename LayoutSettingSelector<NDimSpatial>::DELayout;
template <ck::index_t NDimSpatial>
using RLayout = typename LayoutSettingSelector<NDimSpatial>::RLayout;
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
inline void print_help_msg()
{
std::cerr << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
inline bool parse_cmd_args(int argc,
char* argv[],
ck::utils::conv::ConvParam& problem_size,
ExecutionConfig& config)
{
constexpr int num_execution_config_args =
3; // arguments for do_verification, init_method, time_kernel
constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
constexpr int threshold_to_catch_all_args =
threshold_to_catch_partial_args + num_conv_param_leading_args;
if(argc == 1)
{
// use default
}
// catch only ExecutionConfig arguments
else if(argc == threshold_to_catch_partial_args)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
// catch both ExecutionConfig & ConvParam arguments
else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
problem_size = ck::utils::conv::parse_conv_param(
num_dim_spatial, threshold_to_catch_partial_args, argv);
}
else
{
print_help_msg();
return false;
}
return true;
}
inline HostTensorDescriptor
make_r0_host_tensor_descriptor(const ck::utils::conv::ConvParam& problem_size)
{
std::vector<ck::index_t> dimensions{problem_size.G_, problem_size.N_};
std::copy(begin(problem_size.output_spatial_lengths_),
end(problem_size.output_spatial_lengths_),
std::back_inserter(dimensions));
return HostTensorDescriptor(dimensions);
}
template <typename Lengths, typename Strides>
void unpack_host_tensor_descriptor(const HostTensorDescriptor& descriptor,
Lengths& lengths,
Strides& strides)
{
assert(size(descriptor.GetLengths()) == size(lengths));
std::copy_n(begin(descriptor.GetLengths()), size(descriptor.GetLengths()), begin(lengths));
assert(size(descriptor.GetStrides()) == size(strides));
std::copy_n(begin(descriptor.GetStrides()), size(descriptor.GetStrides()), begin(strides));
}
template <typename Range, typename OutputIterator>
auto copy(const Range& range, OutputIterator iter)
-> decltype(std::copy(std::begin(range), std::end(range), iter))
{
return std::copy(std::begin(range), std::end(range), iter);
}
example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using ADataType = BF16;
using BDataType = BF16;
using AccDataType = FP32;
using CShuffleDataType = FP32;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using ReduceAccDataType = FP32;
using R0DataType = FP32;
using RsDataType = ck::Tuple<R0DataType>;
#include "run_convnd_fwd_max_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using ADataType = FP16;
using BDataType = FP16;
using AccDataType = FP32;
using CShuffleDataType = FP32;
using DsDataType = ck::Tuple<>;
using EDataType = FP16;
using ReduceAccDataType = FP32;
using R0DataType = FP32;
using RsDataType = ck::Tuple<R0DataType>;
#include "run_convnd_fwd_max_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using ADataType = FP32;
using BDataType = FP32;
using AccDataType = FP32;
using CShuffleDataType = FP32;
using DsDataType = ck::Tuple<>;
using EDataType = FP32;
using ReduceAccDataType = FP32;
using R0DataType = FP32;
using RsDataType = ck::Tuple<R0DataType>;
#include "run_convnd_fwd_max_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
#error Should compile this file with ck::int4_t support
#endif
#define BUILD_INT4_EXAMPLE
#include "common.hpp"
using ADataType = I4;
using BDataType = I4;
using KernelADataType = I8;
using KernelBDataType = I8;
using AccDataType = I32;
using CShuffleDataType = I32;
using DsDataType = ck::Tuple<>;
using EDataType = I32;
using ReduceAccDataType = I32;
using R0DataType = I32;
using RsDataType = ck::Tuple<R0DataType>;
#include "run_convnd_fwd_max_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using ADataType = I8;
using BDataType = I8;
using AccDataType = I32;
using CShuffleDataType = I32;
using DsDataType = ck::Tuple<>;
using EDataType = I32;
using ReduceAccDataType = I32;
using R0DataType = I32;
using RsDataType = ck::Tuple<R0DataType>;
#include "run_convnd_fwd_max_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
using QsElementOp = ck::Tuple<PassThrough>;
using RsElementOp = ck::Tuple<PassThrough>;
// ReduceOp
using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
using RsGlobalReduceOp =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
template <ck::index_t NDimSpatial>
using DeviceInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
//######| NDimSpatial| ALayout| BLayout| DELayout| RLayout| AData| BData| AccData| CShuffle| DsData| EData| ReduceAccData| RsData| A| B| CDE| Qs| Rs| Thread| Global| Conv| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CDRThreadTransfer| CDE| RThreadTransfer|
//######| | | | | | Type| Type| Type| DataType| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Fwd|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
//######| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| Operation| Operation| Specialization| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _MPerBlock_NPerBlock| ScalarPerVector| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NPerBlock| |
#ifdef BUILD_INT4_EXAMPLE
< NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>, KernelADataType, KernelBDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType, AElementOp, BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp, ConvSpec, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<64, 4>, 4, 1>;
#else
< NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType, AElementOp, BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp, ConvSpec, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<64, 4>, 4, 1>;
#endif
template <ck::index_t NDimSpatial>
using HostInstance = ck::tensor_operation::host::ReferenceConvFwd
<NDimSpatial, ADataType, BDataType, EDataType, AElementOp, BElementOp, PassThrough>;
// clang-format on
template <ck::index_t NDimSpatial>
bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
const ExecutionConfig& config)
{
static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
#endif
const auto conv_input_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ALayout<NDimSpatial>>(
problem_size);
const auto conv_weight_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<BLayout<NDimSpatial>>(
problem_size);
const auto conv_output_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<DELayout<NDimSpatial>>(
problem_size);
const auto r0_desc = make_r0_host_tensor_descriptor(problem_size);
Tensor<ADataType> conv_input(conv_input_g_n_c_wis_desc);
Tensor<BDataType> conv_weight(conv_weight_g_k_c_xs_desc);
Tensor<EDataType> conv_output_device(conv_output_g_n_k_wos_desc);
Tensor<R0DataType> r0_device(r0_desc);
switch(config.init_method)
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input.begin(),
conv_input.end());
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-8, 7}(conv_weight.begin(),
conv_weight.end());
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-5, 5}(conv_input.begin(), conv_input.end());
ck::utils::FillUniformDistribution<BDataType>{-5, 5}(conv_weight.begin(),
conv_weight.end());
}
DeviceMem conv_input_device_buf(sizeof(ADataType) * conv_input.mDesc.GetElementSpaceSize());
DeviceMem conv_weight_device_buf(sizeof(BDataType) * conv_weight.mDesc.GetElementSpaceSize());
DeviceMem conv_output_device_buf(sizeof(EDataType) *
conv_output_device.mDesc.GetElementSpaceSize());
DeviceMem r0_device_buf(sizeof(R0DataType) * r0_device.mDesc.GetElementSpaceSize());
#ifdef BUILD_INT4_EXAMPLE
const Tensor<KernelADataType> conv_input_converted(conv_input);
const Tensor<KernelBDataType> conv_weight_converted(conv_weight);
conv_input_device_buf.ToDevice(conv_input_converted.mData.data());
conv_weight_device_buf.ToDevice(conv_weight_converted.mData.data());
#else
conv_input_device_buf.ToDevice(conv_input.mData.data());
conv_weight_device_buf.ToDevice(conv_weight.mData.data());
#endif
std::array<ck::index_t, NDimSpatial + 3> conv_input_g_n_c_wis_lengths{},
conv_input_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> conv_weight_g_k_c_xs_lengths{},
conv_weight_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> conv_output_g_n_k_wos_lengths{},
conv_output_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial + 2> r0_lengths{}, r0_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{}, conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{}, input_right_pads{};
unpack_host_tensor_descriptor(
conv_input_g_n_c_wis_desc, conv_input_g_n_c_wis_lengths, conv_input_g_n_c_wis_strides);
unpack_host_tensor_descriptor(
conv_weight_g_k_c_xs_desc, conv_weight_g_k_c_xs_lengths, conv_weight_g_k_c_xs_strides);
unpack_host_tensor_descriptor(
conv_output_g_n_k_wos_desc, conv_output_g_n_k_wos_lengths, conv_output_g_n_k_wos_strides);
unpack_host_tensor_descriptor(r0_desc, r0_lengths, r0_strides);
copy(problem_size.conv_filter_strides_, begin(conv_filter_strides));
copy(problem_size.conv_filter_dilations_, begin(conv_filter_dilations));
copy(problem_size.input_left_pads_, begin(input_left_pads));
copy(problem_size.input_right_pads_, begin(input_right_pads));
// run Conv + Reduction on device
auto conv = DeviceInstance<NDimSpatial>{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(conv_input_device_buf.GetDeviceBuffer(),
conv_weight_device_buf.GetDeviceBuffer(),
std::array<const void*, 0>{},
conv_output_device_buf.GetDeviceBuffer(),
{r0_device_buf.GetDeviceBuffer()},
conv_input_g_n_c_wis_lengths,
conv_input_g_n_c_wis_strides,
conv_weight_g_k_c_xs_lengths,
conv_weight_g_k_c_xs_strides,
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
conv_output_g_n_k_wos_lengths,
conv_output_g_n_k_wos_strides,
r0_lengths,
r0_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
AElementOp{},
BElementOp{},
CDEElementOp{},
QsElementOp{},
RsElementOp{});
if(!conv.IsSupportedArgument(argument))
{
std::cerr << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return false;
}
const float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
const std::size_t flop = problem_size.GetFlops();
const std::size_t num_btype = problem_size.GetByte<ADataType, BDataType, EDataType>();
const float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
const float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< conv.GetTypeString() << std::endl;
if(config.do_verification)
{
Tensor<EDataType> conv_output_host(conv_output_g_n_k_wos_desc);
// run Conv + Reduction on host
auto ref_conv = HostInstance<NDimSpatial>{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(conv_input,
conv_weight,
conv_output_host,
problem_size.conv_filter_strides_,
problem_size.conv_filter_dilations_,
problem_size.input_left_pads_,
problem_size.input_right_pads_,
AElementOp{},
BElementOp{},
PassThrough{});
ref_invoker.Run(ref_argument);
Tensor<R0DataType> r0_host(r0_device.mDesc);
auto reduce0_op = RsThreadReduceOp{}[ck::Number<0>{}];
auto& output_dims = conv_output_g_n_k_wos_desc.GetLengths();
if constexpr(NDimSpatial == 1)
{
for(std::size_t g = 0; g < output_dims[0]; ++g)
{
for(std::size_t n = 0; n < output_dims[1]; ++n)
{
for(std::size_t w = 0; w < output_dims[3]; ++w)
{
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
for(std::size_t k = 0; k < output_dims[2]; ++k)
{
auto e_val =
ck::type_convert<ReduceAccDataType>(conv_output_host(g, n, k, w));
reduce0_op(reduce0_acc, e_val);
}
r0_host(g, n, w) = ck::type_convert<R0DataType>(reduce0_acc);
}
}
}
}
else if constexpr(NDimSpatial == 2)
{
for(std::size_t g = 0; g < output_dims[0]; ++g)
{
for(std::size_t n = 0; n < output_dims[1]; ++n)
{
for(std::size_t h = 0; h < output_dims[3]; ++h)
{
for(std::size_t w = 0; w < output_dims[4]; ++w)
{
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
for(std::size_t k = 0; k < output_dims[2]; ++k)
{
auto e_val = ck::type_convert<ReduceAccDataType>(
conv_output_host(g, n, k, h, w));
reduce0_op(reduce0_acc, e_val);
}
r0_host(g, n, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
}
}
}
}
}
else if constexpr(NDimSpatial == 3)
{
for(std::size_t g = 0; g < output_dims[0]; ++g)
{
for(std::size_t n = 0; n < output_dims[1]; ++n)
{
for(std::size_t d = 0; d < output_dims[3]; ++d)
{
for(std::size_t h = 0; h < output_dims[4]; ++h)
{
for(std::size_t w = 0; w < output_dims[5]; ++w)
{
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
for(std::size_t k = 0; k < output_dims[2]; ++k)
{
auto e_val = ck::type_convert<ReduceAccDataType>(
conv_output_host(g, n, k, d, h, w));
reduce0_op(reduce0_acc, e_val);
}
r0_host(g, n, d, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
}
}
}
}
}
}
conv_output_device_buf.FromDevice(conv_output_device.mData.data());
r0_device_buf.FromDevice(r0_device.mData.data());
return ck::utils::check_err(conv_output_device.mData,
conv_output_host.mData,
"Error: incorrect results! (Matrix E)",
1e-5f,
1e-4f) &&
ck::utils::check_err(r0_device.mData,
r0_host.mData,
"Error: incorrect results! (Matrix R0)",
1e-5f,
1e-4f);
}
return true;
}
bool run_convnd_fwd_max_example(int argc, char* argv[])
{
ck::utils::conv::ConvParam problem_size{
2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ExecutionConfig config;
if(!parse_cmd_args(argc, argv, problem_size, config))
{
return false;
}
switch(problem_size.num_dim_spatial_)
{
case 1: return run_convnd_fwd_max<1>(problem_size, config);
case 2: return run_convnd_fwd_max<2>(problem_size, config);
case 3: return run_convnd_fwd_max<3>(problem_size, config);
}
return false;
}
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View file @ 2732d06c
add_custom_target(example_gemm_reduce_xdl)
add_custom_target(example_gemm_reduce_xdl_max)
add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
add_dependencies(example_gemm_reduce_xdl_max
example_gemm_max_xdl_bf16
example_gemm_max_xdl_fp16
example_gemm_max_xdl_fp32
example_gemm_max_xdl_int8)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare
example_gemm_mean_meansquare_xdl_fp16
example_gemm_mean_meansquare_xdl_fp32
example_gemm_mean_meansquare_xdl_bf16
example_gemm_add_addsquare_xdl_int8)
add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
add_dependencies(example_gemm_reduce_xdl
example_gemm_reduce_xdl_mean_meansquare
example_gemm_reduce_xdl_max
example_gemm_add_add_mean_meansquare_xdl)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
endif()
example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp 0 → 100644
View file @ 2732d06c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_reduce_xdl_common.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
// DataType
using ADataType = INT8;
using BDataType = INT8;
using GemmAccDataType = INT32;
using CShuffleDataType = INT32;
using DsDataType = ck::Tuple<>;
using EDataType = INT8;
using ReduceAccDataType = INT32;
using R0DataType = INT32;
using R1DataType = INT32;
using RsDataType = ck::Tuple<R0DataType, R1DataType>;
// Layout
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
// Elementwise op
using Square = ck::tensor_operation::element_wise::UnarySquare;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
using QsElementOp = ck::Tuple<PassThrough, Square>;
using RsElementOp = ck::Tuple<PassThrough, PassThrough>;
// ReduceOp
using R0ThreadReduceOp = ck::reduce::Add;
using R1ThreadReduceOp = ck::reduce::Add;
using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
<ALayout, // ALayout
BLayout, // BLayout
ELayout, // ELayout
ADataType, // ADataType
BDataType, // BDataType
GemmAccDataType, // GemmAccDataType
CShuffleDataType, // CShuffleDataType
DsDataType, // DsDataType
EDataType, // EDataType
ReduceAccDataType, // ReduceAccDataType
RsDataType, // RsDataType
AElementOp, // AElementwiseOperation
BElementOp, // BElementwiseOperation
CDEElementOp, // CDE ElementwiseOperation
QsElementOp, // Qs Elementwise Operation
RsElementOp, // Rs Elementwise Operation
RsThreadReduceOp, // Thread Reduce Operation
RsGlobalReduceOp, // Global Reduce Operation
GemmDefault, // GEMM Specialization
1, // NumGemmKPrefetchStage
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransfer ThreadCluster Lengths_K0_M_K1
S<1, 0, 2>, // ABlockTransfer ThreadCluster ArrangeOrder
S<1, 0, 2>, // ABlockTransfer SrcAccessOrder
2, // ABlockTransfer SrcVectorDim
16, // ABlockTransfer SrcScalarPerVector
16, // ABlockTransfer DstScalarPerVector_K1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransfer ThreadCluster Lengths_K0_N_K1
S<1, 0, 2>, // BBlockTransfer ThreadCluster ArrangeOrder
S<1, 0, 2>, // BBlockTransfer SrcAccessOrder
2, // BBlockTransfer SrcVectorDim
16, // BBlockTransfer SrcScalarPerVector
16, // BBlockTransfer DstScalarPerVector_K1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<64, 4>, // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
4, // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
1>; // RThread DstScalarPerVector _MPerBlock
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
ReduceAccDataType,
GemmAccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using namespace ck::literals;
template <typename ADataType,
typename BDataType,
typename EDataType,
typename R0DataType,
typename R1DataType,
typename ALayout,
typename BLayout,
typename ELayout,
typename AElementOp,
typename BElementOp,
typename CDEElementOp,
typename QsElementOp,
typename RsElementOp,
typename RsThreadReduceOp,
typename ReduceAccDataType,
typename DeviceOpInstance,
typename ReferenceGemmInstance>
bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t StrideA,
ck::index_t StrideB,
ck::index_t StrideE,
bool do_verification,
int init_method,
bool time_kernel)
{
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor({len}, {stride});
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
switch(init_method)
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(),
a_m_k.end());
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
b_k_n.end());
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end());
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end());
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto qs_element_op = QsElementOp{};
auto rs_element_op = RsElementOp{};
// Prepare GEMM, add, add_square
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
{r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op,
qs_element_op,
rs_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
// init reducetion buffer to 0
r0_device_buf.SetZero();
r1_device_buf.SetZero();
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(do_verification)
{
auto I0 = ck::Number<0>{};
auto I1 = ck::Number<1>{};
Tensor<ReduceAccDataType> e_m_n_host(e_m_n.mDesc);
Tensor<R0DataType> r0_m_host(r0_m.mDesc);
Tensor<R1DataType> r1_m_host(r1_m.mDesc);
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
auto reduce0_op = RsThreadReduceOp{}[I0];
auto reduce1_op = RsThreadReduceOp{}[I1];
for(int m = 0; m < M; ++m)
{
auto reduce0_acc = reduce0_op.template GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.template GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType square_e_val;
auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n_host(m, n));
qs_element_op[I1](square_e_val, e_val);
reduce0_op(reduce0_acc, e_val);
reduce1_op(reduce1_acc, square_e_val);
}
r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
}
e_device_buf.FromDevice(e_m_n.mData.data());
Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host_converted.mData, "Error: Incorrect results c", 1e-2, 1e-2);
r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data());
pass &= ck::utils::check_err(
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
if(pass)
{
std::cout << "Success!" << std::endl;
}
}
if(time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = 2_uz * M * N * K + 3_uz * M * N;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
sizeof(R1DataType) * M;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
<< " GB/s, " << std::endl;
}
return pass;
}
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1152;
ck::index_t K = 512;
ck::index_t StrideA = 512;
ck::index_t StrideB = 512;
ck::index_t StrideE = 1152;
if(argc == 1)
{
// do nothing
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideE = std::stoi(argv[9]);
}
else
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
<< " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
<< std::endl;
exit(EXIT_SUCCESS);
}
return !run_gemm_reduce_add_addsquare_xdl<ADataType,
BDataType,
EDataType,
R0DataType,
R1DataType,
ALayout,
BLayout,
ELayout,
AElementOp,
BElementOp,
CDEElementOp,
QsElementOp,
RsElementOp,
RsThreadReduceOp,
ReduceAccDataType,
DeviceOpInstance,
ReferenceGemmInstance>(
M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
}
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