p85971406
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Delete gpu_collective so for cloud side.

This commit is contained in:
ZPaC 2022-11-11 22:36:29 +08:00
parent 9887716115
commit b7df5799b2
15 changed files with 28 additions and 408 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
cmake/package.cmake
View File

@ -176,13 +176,6 @@ if(ENABLE_CPU)
endif()
if(ENABLE_MPI)
if(ENABLE_GPU)
install(
TARGETS _ms_mpi
DESTINATION ${INSTALL_BASE_DIR}
COMPONENT mindspore
)
endif()
if(ENABLE_CPU)
install(
TARGETS mpi_adapter
@ -206,11 +199,6 @@ endif()
if(ENABLE_GPU)
if(ENABLE_MPI)
install(
TARGETS gpu_collective
DESTINATION ${INSTALL_LIB_DIR}
COMPONENT mindspore
)
install(
TARGETS nvidia_collective
DESTINATION ${INSTALL_LIB_DIR}

14
cmake/package_mac.cmake
View File

@ -133,13 +133,6 @@ if(ENABLE_CPU)
endif()
if(ENABLE_MPI)
if(ENABLE_GPU)
install(
TARGETS _ms_mpi
DESTINATION ${INSTALL_BASE_DIR}
COMPONENT mindspore
)
endif()
if(ENABLE_CPU)
install(
TARGETS mpi_adapter
@ -150,13 +143,6 @@ if(ENABLE_MPI)
endif()
if(ENABLE_GPU)
if(ENABLE_MPI)
install(
TARGETS gpu_collective
DESTINATION ${INSTALL_LIB_DIR}
COMPONENT mindspore
)
endif()
install(
TARGETS gpu_queue
DESTINATION ${INSTALL_LIB_DIR}

12
cmake/package_win.cmake
View File

@ -185,13 +185,6 @@ if(ENABLE_CPU)
endif()
if(ENABLE_MPI)
if(ENABLE_GPU)
install(
TARGETS _ms_mpi
DESTINATION ${INSTALL_BASE_DIR}
COMPONENT mindspore
)
endif()
if(ENABLE_CPU)
install(
TARGETS mpi_adapter
@ -208,11 +201,6 @@ endif()
if(ENABLE_GPU)
if(ENABLE_MPI)
install(
TARGETS gpu_collective
DESTINATION ${INSTALL_LIB_DIR}
COMPONENT mindspore
)
install(
TARGETS nvidia_collective
DESTINATION ${INSTALL_LIB_DIR}

10
mindspore/ccsrc/pipeline/jit/init.cc
View File

@ -32,7 +32,6 @@
#include "include/common/utils/parallel_context.h"
#include "frontend/parallel/costmodel_context.h"
#include "frontend/optimizer/ad/bprop_utils.h"
#include "runtime/collective/gpu_collective_init.h"
#if ((defined ENABLE_CPU) && (!defined _WIN32))
#include "ps/util.h"
#endif
@ -455,15 +454,6 @@ PYBIND11_MODULE(_c_expression, m) {
.def(py::init())
.def_static("reg_op", &OpLib::RegOp, "Register op info.");
(void)m.def("init_gpu_collective", &mindspore::device::gpu::GPUCollectiveInitializer::InitCollective,
"Init gpu collective communication mode.");
(void)m.def("finalize_gpu_collective", &mindspore::device::gpu::GPUCollectiveInitializer::FinalizeCollective,
"Finalize gpu collective communication mode.");
(void)m.def("get_rank_id", &mindspore::device::gpu::GPUCollectiveInitializer::GetRankID,
"Finalize gpu collective communication mode.");
(void)m.def("get_rank_size", &mindspore::device::gpu::GPUCollectiveInitializer::GetRankSize,
"Finalize gpu collective communication mode.");
(void)py::class_<CollectiveManager, std::shared_ptr<CollectiveManager>>(m, "CollectiveManager")
.def_static("get_instance", &CollectiveManager::instance, "Get collective manager instance.")
.def("create_group", &CollectiveManager::CreateCommunicationGroup, "Create collective group.")

2
mindspore/ccsrc/plugin/device/gpu/CMakeLists.txt
View File

@ -90,8 +90,6 @@ endif()
if(ENABLE_GPU)
if(ENABLE_MPI)
set_target_properties(_ms_mpi PROPERTIES INSTALL_RPATH ${MINDSPORE_RPATH})
set_target_properties(nvidia_collective PROPERTIES INSTALL_RPATH ${ORIGIN_PATH})
set_target_properties(gpu_collective PROPERTIES INSTALL_RPATH ${ORIGIN_PATH})
endif()
endif()

23
mindspore/ccsrc/plugin/device/gpu/hal/device/CMakeLists.txt
View File

@ -10,7 +10,6 @@ if(ENABLE_GPU)
list(APPEND DEVICE_SRC_LIST ${CMAKE_SOURCE_DIR}/mindspore/ccsrc/runtime/data_queue/data_queue.h)
file(GLOB_RECURSE DEVICE_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} "*.cc")
list(REMOVE_ITEM DEVICE_SRC_LIST
"mpi/mpi_initializer.cc"
"distribution/collective_wrapper.cc"
"distribution/mpi_wrapper.cc"
"distribution/nccl_wrapper.cc"
@ -29,13 +28,6 @@ if(ENABLE_GPU)
endif()
if(ENABLE_GPU)
if(ENABLE_MPI)
set_property(SOURCE "mpi/mpi_initializer.cc"
PROPERTY COMPILE_DEFINITIONS SUBMODULE_ID=mindspore::SubModuleId::SM_DEVICE)
pybind11_add_module(_ms_mpi NO_EXTRAS "mpi/mpi_initializer.cc")
target_link_libraries(_ms_mpi PRIVATE mindspore::pybind11_module mindspore::ompi)
endif()
file(GLOB_RECURSE CUDA_SRC_LIST RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} "*.cu")
set_property(SOURCE ${CUDA_SRC_LIST} PROPERTY COMPILE_DEFINITIONS SUBMODULE_ID=mindspore::SubModuleId::SM_DEVICE)
@ -47,21 +39,6 @@ if(ENABLE_GPU)
list(REMOVE_DUPLICATES CUDA_NVCC_FLAGS)
cuda_add_library(gpu_hash_table STATIC ${CUDA_SRC_LIST})
endif()
set(GPU_COLLECTIVE_SRCS "distribution/collective_wrapper.cc"
"distribution/mpi_wrapper.cc"
"distribution/nccl_wrapper.cc")
if(ENABLE_MPI)
include(ExternalProject)
# gpu_collective
set_property(SOURCE ${GPU_COLLECTIVE_SRCS}
PROPERTY COMPILE_DEFINITIONS SUBMODULE_ID=mindspore::SubModuleId::SM_DEVICE)
add_library(gpu_collective SHARED ${GPU_COLLECTIVE_SRCS})
target_link_libraries(gpu_collective PRIVATE mindspore::ompi mindspore::nccl)
target_link_libraries(_ms_mpi PRIVATE gpu_collective)
endif()
endif()
set_property(SOURCE ${DEVICE_SRC_LIST} PROPERTY COMPILE_DEFINITIONS SUBMODULE_ID=mindspore::SubModuleId::SM_DEVICE)

44
mindspore/ccsrc/plugin/device/gpu/hal/device/mpi/mpi_initializer.cc
View File

@ -1,44 +0,0 @@
/**
* Copyright 2019 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "plugin/device/gpu/hal/device/mpi/mpi_initializer.h"
#include <dlfcn.h>
#include <mpi.h>
#include <pybind11/operators.h>
#include <iostream>
#include <vector>
#include <string>
namespace mindspore {
namespace device {
namespace gpu {
MPIInitializer &MPIInitializer::GetInstance() {
static MPIInitializer instance;
return instance;
}
int MPIInitializer::get_rank_id(const std::string &group) { return GetRankIDByGroup(group); }
int MPIInitializer::get_rank_size(const std::string &group) { return GetGroupSize(group); }
PYBIND11_MODULE(_ms_mpi, mpi_initializer) {
mpi_initializer.doc() = "mindspore mpi python wrapper";
mpi_initializer.def("get_rank_id", &MPIInitializer::get_rank_id, "get rank id");
mpi_initializer.def("get_rank_size", &MPIInitializer::get_rank_size, "get rank size");
}
} // namespace gpu
} // namespace device
} // namespace mindspore

42
mindspore/ccsrc/plugin/device/gpu/hal/device/mpi/mpi_initializer.h
View File

@ -1,42 +0,0 @@
/**
* Copyright 2019 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef MINDSPORE_CCSRC_RUNTIME_DEVICE_GPU_MPI_MPI_INITIALIZER_H_
#define MINDSPORE_CCSRC_RUNTIME_DEVICE_GPU_MPI_MPI_INITIALIZER_H_
#include <string>
#include "plugin/device/gpu/hal/device/distribution/collective_wrapper.h"
namespace mindspore {
namespace device {
namespace gpu {
class MPIInitializer {
public:
MPIInitializer(MPIInitializer const &) = delete;
MPIInitializer &operator=(const MPIInitializer &) = delete;
static MPIInitializer &GetInstance();
static int get_rank_id(const std::string &group);
static int get_rank_size(const std::string &groups);
private:
MPIInitializer() = default;
~MPIInitializer() = default;
};
} // namespace gpu
} // namespace device
} // namespace mindspore
#endif // MINDSPORE_CCSRC_RUNTIME_DEVICE_GPU_MPI_MPI_INITIALIZER_H_

13
mindspore/ccsrc/plugin/device/gpu/hal/hardware/gpu_deprecated_interface.cc
View File

@ -18,7 +18,6 @@
#include <cuda.h>
#include <vector>
#include <string>
#include "plugin/device/gpu/hal/device/distribution/collective_init.h"
namespace mindspore {
namespace device {
@ -61,18 +60,6 @@ int GPUDeprecatedInterface::GetGPUMultiProcessorCount() {
}
return sm_count;
}
void GPUDeprecatedInterface::GPUInitCollective() { CollectiveInitializer::InitCollective(); }
void GPUDeprecatedInterface::GPUFinalizeCollective() { CollectiveInitializer::FinalizeCollective(); }
uint32_t GPUDeprecatedInterface::GPUGetRankID(const std::string &group_name) {
return CollectiveInitializer::GetRankID(group_name);
}
uint32_t GPUDeprecatedInterface::GPUGetRankSize(const std::string &group_name) {
return CollectiveInitializer::GetRankSize(group_name);
}
} // namespace gpu
} // namespace device
} // namespace mindspore

4
mindspore/ccsrc/plugin/device/gpu/hal/hardware/gpu_deprecated_interface.h
View File

@ -35,10 +35,6 @@ class GPUDeprecatedInterface : public DeprecatedInterface {
int GetGPUCapabilityMajor() override;
int GetGPUCapabilityMinor() override;
int GetGPUMultiProcessorCount() override;
void GPUInitCollective() override;
void GPUFinalizeCollective() override;
uint32_t GPUGetRankID(const std::string &group_name) override;
uint32_t GPUGetRankSize(const std::string &group_name) override;
};
} // namespace gpu
} // namespace device

125
mindspore/ccsrc/plugin/device/gpu/kernel/nccl/nccl_gpu_kernel.cc
View File

@ -20,14 +20,8 @@
namespace mindspore {
namespace kernel {
void NcclGpuKernelMod::SelectCollectiveHandle() {
use_mpi_ = false;
if (use_mpi_) {
collective_handle_ = device::gpu::CollectiveInitializer::instance().collective_handle();
MS_EXCEPTION_IF_NULL(collective_handle_);
} else {
if (!LoadNvidiaCommLib()) {
MS_LOG(EXCEPTION) << "Failed to load nivdia communication library.";
}
if (!LoadNvidiaCommLib()) {
MS_LOG(EXCEPTION) << "Failed to load nivdia communication library.";
}
}
@ -46,128 +40,65 @@ bool NcclGpuKernelMod::LoadNvidiaCommLib() {
bool NcclGpuKernelMod::AllReduce(const void *input_addr, void *output_addr, size_t count, ncclDataType_t data_type,
ncclRedOp_t reduce_op, cudaStream_t stream, const std::string &group_name) {
if (use_mpi_) {
auto all_reduce_funcptr =
reinterpret_cast<kernel::AllReduce>(dlsym(const_cast<void *>(collective_handle_), "AllReduce"));
MS_EXCEPTION_IF_NULL(all_reduce_funcptr);
CHECK_NCCL_RET_WITH_EXCEPT(
kernel_node_, (*all_reduce_funcptr)(input_addr, output_addr, count, data_type, reduce_op, stream, group_name),
"ncclAllReduce failed");
} else {
auto allreduce_func = DlsymFuncObj(AllReduce, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_,
allreduce_func(input_addr, output_addr, count, data_type, reduce_op, group_name, stream),
"ncclAllReduce failed");
}
auto allreduce_func = DlsymFuncObj(AllReduce, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_,
allreduce_func(input_addr, output_addr, count, data_type, reduce_op, group_name, stream),
"ncclAllReduce failed");
return true;
}
bool NcclGpuKernelMod::AllGather(const void *input_addr, void *output_addr, size_t count, ncclDataType_t data_type,
cudaStream_t stream, const std::string &group_name) {
if (use_mpi_) {
auto all_gather_funcptr =
reinterpret_cast<kernel::AllGather>(dlsym(const_cast<void *>(collective_handle_), "AllGather"));
MS_EXCEPTION_IF_NULL(all_gather_funcptr);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_,
(*all_gather_funcptr)(input_addr, output_addr, count, data_type, stream, group_name),
"ncclAllGather failed");
} else {
auto allgather_func = DlsymFuncObj(AllGather, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_,
allgather_func(input_addr, output_addr, count, data_type, group_name, stream),
"ncclAllGather failed");
}
auto allgather_func = DlsymFuncObj(AllGather, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_,
allgather_func(input_addr, output_addr, count, data_type, group_name, stream),
"ncclAllGather failed");
return true;
}
bool NcclGpuKernelMod::ReduceScatter(const void *input_addr, void *output_addr, size_t count, ncclDataType_t data_type,
ncclRedOp_t reduce_op, cudaStream_t stream, const std::string &group_name) {
if (use_mpi_) {
auto reduce_scatter_funcptr =
reinterpret_cast<kernel::ReduceScatter>(dlsym(const_cast<void *>(collective_handle_), "ReduceScatter"));
MS_EXCEPTION_IF_NULL(reduce_scatter_funcptr);
CHECK_NCCL_RET_WITH_EXCEPT(
kernel_node_, (*reduce_scatter_funcptr)(input_addr, output_addr, count, data_type, reduce_op, stream, group_name),
"ncclReduceScatter failed");
} else {
auto reducescatter_func = DlsymFuncObj(ReduceScatter, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(
kernel_node_, reducescatter_func(input_addr, output_addr, count, data_type, reduce_op, group_name, stream),
"ncclReduceScatter failed");
}
auto reducescatter_func = DlsymFuncObj(ReduceScatter, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(
kernel_node_, reducescatter_func(input_addr, output_addr, count, data_type, reduce_op, group_name, stream),
"ncclReduceScatter failed");
return true;
}
bool NcclGpuKernelMod::Broadcast(const void *input_addr, void *output_addr, size_t count, ncclDataType_t data_type,
int root, cudaStream_t stream, const std::string &group_name) {
if (use_mpi_) {
auto broadcast_funcptr =
reinterpret_cast<kernel::Broadcast>(dlsym(const_cast<void *>(collective_handle_), "Broadcast"));
MS_EXCEPTION_IF_NULL(broadcast_funcptr);
CHECK_NCCL_RET_WITH_EXCEPT(
kernel_node_, (*broadcast_funcptr)(input_addr, output_addr, count, data_type, root, stream, group_name),
"ncclBroadcast failed");
} else {
auto broadcast_func = DlsymFuncObj(Broadcast, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_,
broadcast_func(input_addr, output_addr, count, data_type, root, group_name, stream),
"ncclBroadcast failed");
}
auto broadcast_func = DlsymFuncObj(Broadcast, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_,
broadcast_func(input_addr, output_addr, count, data_type, root, group_name, stream),
"ncclBroadcast failed");
return true;
}
bool NcclGpuKernelMod::Send(const void *send_addr, size_t count, ncclDataType_t data_type, int peer_rank,
cudaStream_t stream, const std::string &group_name) {
if (use_mpi_) {
auto nccl_send_func = reinterpret_cast<kernel::Send>(dlsym(const_cast<void *>(collective_handle_), "Send"));
MS_EXCEPTION_IF_NULL(nccl_send_func);
CHECK_NCCL_RET_WITH_EXCEPT(
kernel_node_, (*nccl_send_func)(send_addr, count, data_type, peer_rank, stream, group_name), "ncclSend failed");
} else {
auto send_func = DlsymFuncObj(Send, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, send_func(send_addr, count, data_type, peer_rank, group_name, stream),
"ncclSend failed");
}
auto send_func = DlsymFuncObj(Send, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, send_func(send_addr, count, data_type, peer_rank, group_name, stream),
"ncclSend failed");
return true;
}
bool NcclGpuKernelMod::Recv(void *recv_addr, size_t count, ncclDataType_t data_type, int peer_rank, cudaStream_t stream,
const std::string &group_name) {
if (use_mpi_) {
auto nccl_recv_func = reinterpret_cast<kernel::Recv>(dlsym(const_cast<void *>(collective_handle_), "Recv"));
MS_EXCEPTION_IF_NULL(nccl_recv_func);
CHECK_NCCL_RET_WITH_EXCEPT(
kernel_node_, (*nccl_recv_func)(recv_addr, count, data_type, peer_rank, stream, group_name), "ncclRecv failed");
} else {
auto recv_func = DlsymFuncObj(Recv, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, recv_func(recv_addr, count, data_type, peer_rank, group_name, stream),
"ncclRecv failed");
}
auto recv_func = DlsymFuncObj(Recv, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, recv_func(recv_addr, count, data_type, peer_rank, group_name, stream),
"ncclRecv failed");
return true;
}
bool NcclGpuKernelMod::GroupStart() {
if (use_mpi_) {
auto nccl_gstart_func =
reinterpret_cast<kernel::GroupStart>(dlsym(const_cast<void *>(collective_handle_), "GroupStart"));
MS_EXCEPTION_IF_NULL(nccl_gstart_func);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, (*nccl_gstart_func)(), "ncclGroupStart failed");
} else {
auto groupstart_func = DlsymFuncObj(GroupStart, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, groupstart_func(), "ncclGroupStart failed");
}
auto groupstart_func = DlsymFuncObj(GroupStart, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, groupstart_func(), "ncclGroupStart failed");
return true;
}
bool NcclGpuKernelMod::GroupEnd() {
if (use_mpi_) {
auto nccl_gend_func = reinterpret_cast<kernel::GroupEnd>(dlsym(const_cast<void *>(collective_handle_), "GroupEnd"));
MS_EXCEPTION_IF_NULL(nccl_gend_func);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, (*nccl_gend_func)(), "ncclGroupEnd failed");
} else {
auto groupend_func = DlsymFuncObj(GroupEnd, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, groupend_func(), "ncclGroupEnd failed");
}
auto groupend_func = DlsymFuncObj(GroupEnd, nvidia_collective_handle_);
CHECK_NCCL_RET_WITH_EXCEPT(kernel_node_, groupend_func(), "ncclGroupEnd failed");
return true;
}
} // namespace kernel

1
mindspore/ccsrc/plugin/device/gpu/kernel/nccl/nccl_gpu_kernel.h
View File

@ -24,7 +24,6 @@
#include "plugin/device/gpu/kernel/gpu_kernel.h"
#include "plugin/device/gpu/kernel/gpu_kernel_factory.h"
#include "plugin/device/gpu/kernel/kernel_constants.h"
#include "plugin/device/gpu/hal/device/distribution/collective_init.h"
#include "plugin/device/gpu/hal/hardware/nvidia_collective_comm_lib.h"
#include "runtime/collective/collective_comm_lib_loader.h"

89
mindspore/ccsrc/runtime/collective/gpu_collective_init.cc
View File

@ -1,89 +0,0 @@
/**
* Copyright 2019 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "runtime/collective/gpu_collective_init.h"
#include "runtime/hardware/device_context_manager.h"
#include "utils/log_adapter.h"
namespace mindspore {
namespace device {
namespace gpu {
void GPUCollectiveInitializer::InitCollective() {
auto ms_context = MsContext::GetInstance();
MS_EXCEPTION_IF_NULL(ms_context);
auto device_target = ms_context->get_param<std::string>(MS_CTX_DEVICE_TARGET);
if (device_target != kGPUDevice) {
MS_LOG(EXCEPTION) << "You are trying to call 'init('nccl')', Please check "
"this MindSpore package is GPU version and built with NCCL.";
}
const auto &device_context = device::DeviceContextManager::GetInstance().GetOrCreateDeviceContext(
{kGPUDevice, ms_context->get_param<uint32_t>(MS_CTX_DEVICE_ID)});
MS_EXCEPTION_IF_NULL(device_context);
auto deprecated_ptr = device_context->GetDeprecatedInterface();
MS_EXCEPTION_IF_NULL(deprecated_ptr);
deprecated_ptr->GPUInitCollective();
}
void GPUCollectiveInitializer::FinalizeCollective() {
auto ms_context = MsContext::GetInstance();
MS_EXCEPTION_IF_NULL(ms_context);
auto device_target = ms_context->get_param<std::string>(MS_CTX_DEVICE_TARGET);
if (device_target != kGPUDevice) {
MS_LOG(EXCEPTION) << "You are trying to call 'finalize('nccl')', Please check "
"this MindSpore package is GPU version and built with NCCL.";
}
const auto &device_context = device::DeviceContextManager::GetInstance().GetOrCreateDeviceContext(
{kGPUDevice, ms_context->get_param<uint32_t>(MS_CTX_DEVICE_ID)});
MS_EXCEPTION_IF_NULL(device_context);
auto deprecated_ptr = device_context->GetDeprecatedInterface();
MS_EXCEPTION_IF_NULL(deprecated_ptr);
deprecated_ptr->GPUFinalizeCollective();
}
uint32_t GPUCollectiveInitializer::GetRankID(const std::string &group_name) {
auto ms_context = MsContext::GetInstance();
MS_EXCEPTION_IF_NULL(ms_context);
auto device_target = ms_context->get_param<std::string>(MS_CTX_DEVICE_TARGET);
if (device_target != kGPUDevice) {
MS_LOG(EXCEPTION) << "You are trying to call 'GetRankID', Please check "
"this MindSpore package is GPU version and built with NCCL.";
}
const auto &device_context = device::DeviceContextManager::GetInstance().GetOrCreateDeviceContext(
{kGPUDevice, ms_context->get_param<uint32_t>(MS_CTX_DEVICE_ID)});
MS_EXCEPTION_IF_NULL(device_context);
auto deprecated_ptr = device_context->GetDeprecatedInterface();
MS_EXCEPTION_IF_NULL(deprecated_ptr);
return deprecated_ptr->GPUGetRankID(group_name);
}
uint32_t GPUCollectiveInitializer::GetRankSize(const std::string &group_name) {
auto ms_context = MsContext::GetInstance();
MS_EXCEPTION_IF_NULL(ms_context);
auto device_target = ms_context->get_param<std::string>(MS_CTX_DEVICE_TARGET);
if (device_target != kGPUDevice) {
MS_LOG(EXCEPTION) << "You are trying to call 'GetRankSize', Please check "
"this MindSpore package is GPU version and built with NCCL.";
}
const auto &device_context = device::DeviceContextManager::GetInstance().GetOrCreateDeviceContext(
{kGPUDevice, ms_context->get_param<uint32_t>(MS_CTX_DEVICE_ID)});
MS_EXCEPTION_IF_NULL(device_context);
auto deprecated_ptr = device_context->GetDeprecatedInterface();
MS_EXCEPTION_IF_NULL(deprecated_ptr);
return deprecated_ptr->GPUGetRankSize(group_name);
}
} // namespace gpu
} // namespace device
} // namespace mindspore

41
mindspore/ccsrc/runtime/collective/gpu_collective_init.h
View File

@ -1,41 +0,0 @@
/**
* Copyright 2019 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef MINDSPORE_CCSRC_RUNTIME_COLLECTIVE_GPU_COLLECTIVE_INIT_H_
#define MINDSPORE_CCSRC_RUNTIME_COLLECTIVE_GPU_COLLECTIVE_INIT_H_
#include <string>
#include "include/backend/visible.h"
namespace mindspore {
namespace device {
namespace gpu {
class BACKEND_EXPORT GPUCollectiveInitializer {
public:
GPUCollectiveInitializer() = default;
~GPUCollectiveInitializer() = default;
GPUCollectiveInitializer(GPUCollectiveInitializer const &) = delete;
GPUCollectiveInitializer &operator=(const GPUCollectiveInitializer &) = delete;
static void InitCollective();
static void FinalizeCollective();
static uint32_t GetRankID(const std::string &group_name);
static uint32_t GetRankSize(const std::string &group_name);
};
} // namespace gpu
} // namespace device
} // namespace mindspore
#endif // MINDSPORE_CCSRC_RUNTIME_COLLECTIVE_GPU_COLLECTIVE_INIT_H_

4
mindspore/ccsrc/runtime/hardware/deprecated_interface.h
View File

@ -61,10 +61,6 @@ class DeprecatedInterface {
virtual int GetGPUCapabilityMajor() { return -1; }
virtual int GetGPUCapabilityMinor() { return -1; }
virtual int GetGPUMultiProcessorCount() { return -1; }
virtual void GPUInitCollective() {}
virtual void GPUFinalizeCollective() {}
virtual uint32_t GPUGetRankID(const std::string &group_name) { return 0; }
virtual uint32_t GPUGetRankSize(const std::string &group_name) { return 0; }
};
} // namespace device
} // namespace mindspore