Skip to content

GitLab

Commit 8fb2b172 authored by Jing Zhang's avatar Jing Zhang
Browse files

add grouped gemm profiler

parent 65d67fb7
Show whitespace changes
Inline Side-by-side
Showing with 772 additions and 31 deletions
include/ck/tensor_operation/gpu/device/device_gemm.hpp
View file @ 8fb2b172
......@@ -59,6 +59,12 @@ struct DeviceGemm : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGemmPtr = std::unique_ptr<
DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
......@@ -76,8 +82,8 @@ struct DeviceGroupedGemm : public BaseOperator
template <typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
using DeviceGemmPtr = std::unique_ptr<
DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
using DeviceGroupedGemmPtr = std::unique_ptr<
DeviceGroupedGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp
View file @ 8fb2b172
......@@ -331,19 +331,17 @@ struct DeviceGroupedGemmXdl
{
GemmShape_arg(i) = arg.GemmShape_[i];
std::cout << "arg.a_grid_desc_k0_m_k1_{"
std::cout << "group: " << i << " arg.a_grid_desc_k0_m_k1_{"
<< GemmShape_arg[i].a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
<< GemmShape_arg[i].a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< GemmShape_arg[i].a_grid_desc_k0_m_k1_.GetLength(I2) << "}"
<< std::endl;
<< GemmShape_arg[i].a_grid_desc_k0_m_k1_.GetLength(I2) << "}";
std::cout << "arg.b_grid_desc_k0_n_k1_{"
std::cout << ", arg.b_grid_desc_k0_n_k1_{"
<< GemmShape_arg[i].b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
<< GemmShape_arg[i].b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< GemmShape_arg[i].b_grid_desc_k0_n_k1_.GetLength(I2) << "}"
<< std::endl;
<< GemmShape_arg[i].b_grid_desc_k0_n_k1_.GetLength(I2) << "}";
std::cout << "arg.c_grid_desc_m_n_{ "
std::cout << ", arg.c_grid_desc_m_n_{ "
<< GemmShape_arg[i].c_grid_desc_m_n_.GetLength(I0) << ", "
<< GemmShape_arg[i].c_grid_desc_m_n_.GetLength(I1) << "}"
<< std::endl;
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 8fb2b172
......@@ -28,3 +28,4 @@ add_subdirectory(conv2d_fwd_bias_relu_add)
add_subdirectory(conv2d_fwd_bias_relu_atomic_add)
add_subdirectory(conv2d_bwd_data)
add_subdirectory(reduce)
add_subdirectory(grouped_gemm)
library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt 0 → 100644
View file @ 8fb2b172
# device_grouped_gemm_instance
set(DEVICE_GROUPED_GEMM_INSTANCE_SOURCE
device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp;
#device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp;
#device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp;
#device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp;
)
add_library(device_grouped_gemm_instance SHARED ${DEVICE_GROUPED_GEMM_INSTANCE_SOURCE})
target_compile_features(device_grouped_gemm_instance PUBLIC)
set_target_properties(device_grouped_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
install(TARGETS device_grouped_gemm_instance LIBRARY DESTINATION lib)
clang_tidy_check(device_grouped_gemm_instance)
library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp 0 → 100644
View file @ 8fb2b172
#include <stdlib.h>
#include "config.hpp"
#include "device_grouped_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances =
std::tuple<
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 256, 4, 8, 32, 32, 1, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 64, 4, 8, 32, 32, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 32, 4, 8, 32, 32, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 256, 4, 8, 16, 16, 1, 8, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 128, 4, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 64, 4, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 16, 32, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
//DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 16, 4, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>
DeviceGroupedGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1>
// clang-format on
>;
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGroupedGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances{});
}
} // namespace device_grouped_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ 8fb2b172
......@@ -22,30 +22,33 @@ include_directories(BEFORE
# ck_profiler
set(PROFILER_SOURCE
src/profiler.cpp
src/profile_gemm.cpp
src/profile_gemm_bias_2d.cpp
src/profile_gemm_bias_relu.cpp
src/profile_gemm_bias_relu_add.cpp
src/profile_batched_gemm.cpp
src/profile_conv_fwd.cpp
src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp
src/profile_conv_fwd_bias_relu_atomic_add.cpp
src/profile_conv_bwd_data.cpp
src/profile_reduce.cpp
#src/profile_gemm.cpp
#src/profile_gemm_bias_2d.cpp
#src/profile_gemm_bias_relu.cpp
#src/profile_gemm_bias_relu_add.cpp
#src/profile_batched_gemm.cpp
#src/profile_conv_fwd.cpp
#src/profile_conv_fwd_bias_relu.cpp
#src/profile_conv_fwd_bias_relu_add.cpp
#src/profile_conv_fwd_bias_relu_atomic_add.cpp
#src/profile_conv_bwd_data.cpp
#src/profile_reduce.cpp
src/profile_grouped_gemm.cpp
)
add_executable(ckProfiler ${PROFILER_SOURCE})
target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
#target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
#target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
#target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
#target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance)
#target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
#target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
#target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
#target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
#target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_atomic_add_instance)
#target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_data_instance)
#target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
#target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
profiler/include/profile_grouped_gemm_impl.hpp 0 → 100644
View file @ 8fb2b172
#pragma once
#include <iomanip>
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm.hpp"
#include "reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
using DeviceGroupedGemmNoOpPtr =
ck::tensor_operation::device::DeviceGroupedGemmPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGroupedGemmNoOpPtr>&);
//void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGroupedGemmNoOpPtr>&);
//void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGroupedGemmNoOpPtr>&);
//void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGroupedGemmNoOpPtr>&);
} // namespace device_grouped_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_grouped_gemm_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
std::vector<int> Ms,
std::vector<int> Ns,
std::vector<int> Ks,
std::vector<int> StrideAs,
std::vector<int> StrideBs,
std::vector<int> StrideCs)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
std::vector<Tensor<ADataType>> a_m_k;
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n;
for(int i = 0; i < Ms.size(); i++)
{
a_m_k.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
Ms[i], Ks[i], StrideAs[i], ALayout{})));
b_k_n.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
Ks[i], Ns[i], StrideBs[i], BLayout{})));
c_m_n.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
Ms[i], Ns[i], StrideCs[i], CLayout{})));
std::cout << "a_m_k[" << i << "]:" << a_m_k[i].mDesc << std::endl;
std::cout << "b_k_n[" << i << "]:" << b_k_n[i].mDesc << std::endl;
std::cout << "c_m_n[" << i << "]:" << c_m_n[i].mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0: break;
case 1:
a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
// set zero to c_device_buf
c_m_n[i].GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
// if(do_verification)
// {
// }
std::vector<DeviceMem> a_device_buf, b_device_buf, c_device_buf;
//DeviceMem a_device_buf(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpace());
//DeviceMem b_device_buf(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpace());
//DeviceMem c_device_buf(sizeof(CDataType) * c_m_n[i].mDesc.GetElementSpace());
for(int i = 0; i < Ms.size(); i++)
{
a_device_buf.push_back(DeviceMem(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpace()));
b_device_buf.push_back(DeviceMem(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpace()));
c_device_buf.push_back(DeviceMem(sizeof(CDataType) * c_m_n[i].mDesc.GetElementSpace()));
a_device_buf[i].ToDevice(a_m_k[i].mData.data());
b_device_buf[i].ToDevice(b_k_n[i].mData.data());
c_device_buf[i].ToDevice(c_m_n[i].mData.data());
}
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_grouped_gemm_instance::DeviceGroupedGemmNoOpPtr> gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
#if 0
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
if(KBatch > 1)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
else
{
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_grouped_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
#endif
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
#if 0
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
KBatch);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + sizeof(CDataType) * 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, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value)
{
Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<float> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<float> c_m_n_device_f32_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
bf16_to_f32_(a_m_k, a_f32_m_k);
bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<float, float, float, AElementOp, BElementOp, CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_f32_m_k,
b_f32_k_n,
c_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_f32_result);
if(do_log)
{
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
else
{
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_result);
if(do_log)
{
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
#endif
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/src/profile_grouped_gemm.cpp 0 → 100644
View file @ 8fb2b172
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_grouped_gemm_impl.hpp"
enum GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
std::vector<int> stringToArray(char *input)
{
std::vector<int> out;
std::istringstream in(input);
std::string item;
while (std::getline(in, item, ',')) {
out.push_back(std::stoi(item));
}
return out;
}
int profile_grouped_gemm(int argc, char* argv[])
{
if(!(argc == 14))
{
printf("arg1: tensor operation (grouped_gemm: Grouped GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs\n");
exit(1);
}
const int data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const int layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const int nrepeat = std::stoi(argv[7]);
const auto Ms = stringToArray(argv[8]);
const auto Ns = stringToArray(argv[9]);
const auto Ks = stringToArray(argv[10]);
const auto StrideAs = stringToArray(argv[11]);
const auto StrideBs = stringToArray(argv[12]);
const auto StrideCs = stringToArray(argv[13]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs);
}
#if 0
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_impl<float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<int8_t,
int8_t,
int8_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
#endif
return 1;
}
profiler/src/profiler.cpp
View file @ 8fb2b172
......@@ -15,9 +15,11 @@ int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_fwd_bias_relu_atomic_add(int, char*[]);
int profile_conv_bwd_data(int, char*[]);
int profile_reduce(int, char*[]);
int profile_grouped_gemm(int, char*[]);
int main(int argc, char* argv[])
{
#if 0
if(strcmp(argv[1], "gemm") == 0)
{
return profile_gemm(argc, argv);
......@@ -62,6 +64,10 @@ int main(int argc, char* argv[])
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "grouped_gemm") == 0)
{
return profile_grouped_gemm(argc, argv);
}
else
{
// clang-format off
......@@ -74,9 +80,13 @@ int main(int argc, char* argv[])
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
" conv_bwd: BackwardConvolution\n"
" grouped_gemm: Grouped Gemm\n"
" reduce: REDUCE\n");
// clang-format on
return 0;
}
#else
profile_grouped_gemm(argc, argv);
#endif
}
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