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

upgrade Ascend software suite 29 Apr 22

This commit is contained in:
yanghaoran 2022-05-23 20:51:25 +08:00
parent 8be4f542e7
commit f7dfcb0494
28 changed files with 119 additions and 120 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
graphengine

@ -1 +1 @@
Subproject commit d63e074e5882d0461769cc0893e7e722ff1695f7
Subproject commit 2537ca0608f64be64d6d98b048b7ec4e48f4acc3

9
mindspore/ccsrc/CMakeLists.txt
View File

@ -430,9 +430,12 @@ if(MODE_ASCEND_ALL OR MODE_ASCEND_ACL)
endif()
if(ENABLE_D)
find_library(GE_RUNNER ge_runner ${ASCEND_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH})
find_library(GRAPH graph ${ASCEND_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH})
find_library(HCCL hccl ${ASCEND_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH})
find_library(GE_RUNNER ge_runner ${ASCEND_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(GRAPH graph ${ASCEND_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(HCCL hccl ${ASCEND_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
target_link_libraries(mindspore_backend PUBLIC ${GE_RUNNER} ${GRAPH} ${HCCL})
endif()

20
mindspore/ccsrc/cxx_api/CMakeLists.txt
View File

@ -130,25 +130,25 @@ if(ENABLE_D OR ENABLE_ACL)
# 310 mode
add_compile_definitions(ENABLE_DVPP_INTERFACE)
find_library(acl libascendcl.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(acl_cblas libacl_cblas.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(acl_dvpp libacl_dvpp.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(acl_runtime libruntime.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(ge_compiler libge_compiler.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(libplatform libplatform.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(libcompress libcompress.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(libopskernel libopskernel.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(libaicore_utils libaicore_utils.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(libaicpu_engine_common libaicpu_engine_common.so ${ASCEND_RUNTIME_PATH}
${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
target_link_libraries(mindspore_shared_lib PRIVATE -Wl,--no-as-needed graph ${ge_compiler}
${acl_retr} ${acl_cblas} ${acl_dvpp} ${acl_runtime} ${libplatform} ${libcompress} ${libopskernel}

4
mindspore/ccsrc/plugin/device/ascend/hal/device/CMakeLists.txt
View File

@ -42,7 +42,9 @@ if(ENABLE_D)
set(ASCEND_TOOLKIT_RUNTIME_PATH ${ASCEND_PATH}/ascend-toolkit/latest/fwkacllib/lib64)
set(ASCEND_RUNTIME_PATH ${ASCEND_PATH}/fwkacllib/lib64)
set(ASCEND_CANN_RUNTIME_PATH ${ASCEND_PATH}/latest/fwkacllib/lib64)
find_library(HCCL hccl ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
set(ASCEND_CANN_RUNTIME_NEW_PATH ${ASCEND_PATH}/latest/lib64)
find_library(HCCL hccl ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
file(GLOB_RECURSE ASCEND_COLLECTIVE_LIST "distribute/mpi_collective_group.cc"
"distribute/collective_group_wrapper.cc")
set_property(SOURCE ${ASCEND_COLLECTIVE_LIST}

8
mindspore/ccsrc/plugin/device/ascend/hal/device/ge_runtime/task/aicpu_task.cc
View File

@ -86,11 +86,11 @@ void AicpuTask::Distribute() {
rtArgsEx_t argsInfo = {};
argsInfo.args = args_;
argsInfo.argsSize = args_size;
rt_ret = rtCpuKernelLaunchWithFlagV2(reinterpret_cast<const void *>(task_info_->so_name().data()),
reinterpret_cast<const void *>(task_info_->kernel_name().data()), 1, &argsInfo,
nullptr, stream_, cpu_flag);
rt_ret = rtCpuKernelLaunchWithFlag(reinterpret_cast<const void *>(task_info_->so_name().data()),
reinterpret_cast<const void *>(task_info_->kernel_name().data()), 1, &argsInfo,
nullptr, stream_, cpu_flag);
if (rt_ret != RT_ERROR_NONE) {
MS_LOG(EXCEPTION) << "Call rt api rtCpuKernelLaunchWithFlagV2 failed, ret: " << rt_ret;
MS_LOG(EXCEPTION) << "Call rt api rtCpuKernelLaunchWithFlag failed, ret: " << rt_ret;
}
MS_LOG(INFO) << "Distribute AicpuTask end.";

2
mindspore/ccsrc/plugin/device/ascend/hal/device/ge_runtime/task/tbe_task.cc
View File

@ -92,7 +92,7 @@ void TbeTask::Distribute() {
rtArgsEx_t args_info = {};
args_info.args = args_;
args_info.argsSize = args_size;
rt_ret = rtKernelLaunchWithFlagV2(stub_func_, task_info_->block_dim(), &args_info, nullptr, stream_, dump_flag);
rt_ret = rtKernelLaunchWithFlag(stub_func_, task_info_->block_dim(), &args_info, nullptr, stream_, dump_flag);
if (rt_ret != RT_ERROR_NONE) {
MS_LOG(EXCEPTION) << "Call rt api rtKernelLaunch failed, ret: " << rt_ret << " mem size " << args_size;
}

38
mindspore/ccsrc/plugin/device/ascend/hal/device/profiling/profiling_manager.cc
View File

@ -57,13 +57,8 @@ uint64_t GetProfilingModule() {
}
Status ProfilingManager::PluginInit() const {
if (prof_cb_.msprofReporterCallback == nullptr) {
MS_LOG(ERROR) << "MsprofReporterCallback callback is nullptr.";
return PROF_FAILED;
}
int32_t ret = prof_cb_.msprofReporterCallback(static_cast<uint32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<uint32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_INIT),
nullptr, 0);
int32_t ret = MsprofReportData(static_cast<uint32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<uint32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_INIT), nullptr, 0);
if (ret != UintToInt(PROF_SUCCESS)) {
MS_LOG(ERROR) << "MsprofReporter init failed, ret: " << ret;
return PROF_FAILED;
@ -72,13 +67,9 @@ Status ProfilingManager::PluginInit() const {
}
void ProfilingManager::PluginUnInit() const {
if (prof_cb_.msprofReporterCallback == nullptr) {
MS_LOG(ERROR) << "MsprofReporterCallback callback is nullptr.";
return;
}
int32_t cb_ret = prof_cb_.msprofReporterCallback(
static_cast<uint32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<uint32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_UNINIT), nullptr, 0);
int32_t cb_ret =
MsprofReportData(static_cast<uint32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<uint32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_UNINIT), nullptr, 0);
if (cb_ret != 0) {
MS_LOG(WARNING) << "profiling plugin uninit failed, ret:%d" << cb_ret;
}
@ -130,7 +121,7 @@ bool ProfilingManager::InitProfiling(const std::string &profiling_path, uint32_t
}
bool ProfilingManager::ProfRegisterCtrlCallback() const {
rtError_t rt_ret = rtProfRegisterCtrlCallback(GE, CtrlCallbackHandle);
rtError_t rt_ret = MsprofRegisterCallback(GE, CtrlCallbackHandle);
if (rt_ret != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Call rtProfRegisterCtrlCallback failed.";
return false;
@ -161,14 +152,9 @@ void ProfilingManager::ReportErrorMessage() const {
}
Status ProfilingManager::CallMsprofReport(const NotNull<ReporterData *> reporter_data) const {
if (prof_cb_.msprofReporterCallback == nullptr) {
MS_LOG(ERROR) << "MsprofReporterCallback callback is nullptr.";
return PROF_FAILED;
}
int32_t ret =
prof_cb_.msprofReporterCallback(static_cast<int32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<int32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_REPORT),
static_cast<void *>(reporter_data.get()), sizeof(ReporterData));
int32_t ret = MsprofReportData(static_cast<int32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<int32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_REPORT),
static_cast<void *>(reporter_data.get()), sizeof(ReporterData));
if (ret != UintToInt(PROF_SUCCESS)) {
ReportErrorMessage();
@ -249,9 +235,9 @@ void ProfilingManager::QueryHashId(const int32_t &device_id, const std::string &
hash_data.dataLen = src_str.size();
hash_data.data = reinterpret_cast<unsigned char *>(const_cast<char *>(src_str.c_str()));
const int32_t ret = prof_cb_.msprofReporterCallback(
static_cast<int32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<int32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_HASH), &hash_data, sizeof(MsprofHashData));
const int32_t ret = MsprofReportData(static_cast<int32_t>(MsprofReporterModuleId::MSPROF_MODULE_FRAMEWORK),
static_cast<int32_t>(MsprofReporterCallbackType::MSPROF_REPORTER_HASH),
&hash_data, sizeof(MsprofHashData));
if (ret != UintToInt(PROF_SUCCESS)) {
ReportErrorMessage();
MS_LOG(EXCEPTION) << "[Profiling] Query hash id of long string failed, src string is " << src_str.c_str()

2
mindspore/ccsrc/plugin/device/ascend/hal/device/profiling/profiling_manager.h
View File

@ -38,7 +38,6 @@ namespace device {
namespace ascend {
struct MsprofCallback {
MsprofCtrlCallback msprofCtrlCallback;
MsprofSetDeviceCallback msprofSetDeviceCallback;
MsprofReporterCallback msprofReporterCallback;
};
@ -68,7 +67,6 @@ class ProfilingManager {
const struct MsprofCallback &GetMsprofCallback() const { return prof_cb_; }
void SetMsprofCtrlCallback(MsprofCtrlCallback func) { prof_cb_.msprofCtrlCallback = func; }
void SetMsprofReporterCallback(MsprofReporterCallback func) { prof_cb_.msprofReporterCallback = func; }
void SetMsprofSetDeviceCallback(MsprofSetDeviceCallback func) { prof_cb_.msprofSetDeviceCallback = func; }
Status GetProfConf(NotNull<MsprofGeOptions *> prof);
Status ProfCommandHandle(ProfCommandHandleType type);
Status ProfHandleInit();

17
mindspore/ccsrc/plugin/device/ascend/hal/hccl_adapter/plugin/CMakeLists.txt
View File

@ -6,10 +6,12 @@ set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/nnae/latest/fwkacllib/l
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/ascend-toolkit/latest/fwkacllib/lib64)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/fwkacllib/lib64)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/latest/fwkacllib/lib64)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/latest/lib64)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/nnae/latest/fwkacllib/lib64/plugin/opskernel)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/ascend-toolkit/latest/fwkacllib/lib64/plugin/opskernel)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/fwkacllib/lib64/plugin/opskernel)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/latest/fwkacllib/lib64/plugin/opskernel)
set(MINDSPORE_RPATH ${MINDSPORE_RPATH}:/usr/local/Ascend/latest/lib64/plugin/opskernel)
set_target_properties(hccl_plugin PROPERTIES INSTALL_RPATH ${MINDSPORE_RPATH})
if(DEFINED ENV{D_LINK_PATH})
@ -30,18 +32,23 @@ else()
set(ASCEND_PATH /usr/local/Ascend)
endif()
set(ASCEND_RUNTIME_PATH ${ASCEND_PATH}/fwkacllib/lib64)
set(ASCEND_CANN_RUNTIME_NEW_PATH ${ASCEND_PATH}/latest/lib64)
set(ASCEND_PLUGIN_PATH ${ASCEND_RUNTIME_PATH}/plugin/opskernel)
set(ASCEND_PLUGIN_NEW_PATH ${ASCEND_CANN_RUNTIME_NEW_PATH}/plugin/opskernel)
set(ASCEND_TOOLKIT_RUNTIME_PATH ${ASCEND_PATH}/ascend-toolkit/latest/fwkacllib/lib64)
set(ASCEND_TOOLKIT_PLUGIN_PATH ${ASCEND_TOOLKIT_RUNTIME_PATH}/plugin/opskernel)
set(ASCEND_CANN_RUNTIME_PATH ${ASCEND_PATH}/latest/fwkacllib/lib64)
set(ASCEND_CANN_PLUGIN_PATH ${ASCEND_CANN_RUNTIME_PATH}/plugin/opskernel)
endif()
find_library(HCCL hccl ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
find_library(REGISTER register ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
find_library(HCCL hccl ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(REGISTER register ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(HCCL_ADPTER hcom_graph_adaptor ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_PATH})
find_library(HCCL_RA ra ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
${ASCEND_CANN_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(HCCL_RA ra ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
find_library(HCCL_BUILDER hcom_opskernel_builder ${ASCEND_PLUGIN_PATH} ${ASCEND_TOOLKIT_PLUGIN_PATH}
${ASCEND_CANN_PLUGIN_PATH})
${ASCEND_CANN_PLUGIN_PATH} ${ASCEND_PLUGIN_NEW_PATH})
target_link_libraries(hccl_plugin -Wl,--no-as-needed ${HCCL} ${HCCL_ADPTER} ${REGISTER} ${HCCL_BUILDER} ${HCCL_RA})

6
mindspore/ccsrc/plugin/device/ascend/kernel/aicpu/aicpu_kernel_mod.cc
View File

@ -165,9 +165,9 @@ bool AicpuOpKernelMod::Launch(const std::vector<AddressPtr> &inputs, const std::
rtArgsEx_t argsInfo = {};
argsInfo.args = args_.data();
argsInfo.argsSize = static_cast<uint32_t>(args_.length());
if (rtCpuKernelLaunchWithFlagV2(reinterpret_cast<const void *>(node_so_.c_str()),
reinterpret_cast<const void *>(node_name_.c_str()), 1, &argsInfo, nullptr, stream_,
flag) != RT_ERROR_NONE) {
if (rtCpuKernelLaunchWithFlag(reinterpret_cast<const void *>(node_so_.c_str()),
reinterpret_cast<const void *>(node_name_.c_str()), 1, &argsInfo, nullptr, stream_,
flag) != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Aicpu op launch failed!";
return false;

6
mindspore/ccsrc/plugin/device/ascend/kernel/aicpu/dynamic_aicpu_kernel_mod.cc
View File

@ -164,9 +164,9 @@ bool DynamicAicpuOpKernelMod::Launch(const std::vector<AddressPtr> &inputs, cons
rtArgsEx_t argsInfo = {};
argsInfo.args = args_.data();
argsInfo.argsSize = static_cast<uint32_t>(args_.length());
ret = rtCpuKernelLaunchWithFlagV2(reinterpret_cast<const void *>(node_so_.c_str()),
reinterpret_cast<const void *>(node_name_.c_str()), 1, &argsInfo, nullptr, stream_,
flag);
ret =
rtCpuKernelLaunchWithFlag(reinterpret_cast<const void *>(node_so_.c_str()),
reinterpret_cast<const void *>(node_name_.c_str()), 1, &argsInfo, nullptr, stream_, flag);
if (ret != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Aicpu op launch failed!";
return false;

7
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/dynamic_tbe_kernel_mod.cc
View File

@ -279,17 +279,16 @@ bool DynamicTbeKernelMod::Launch(const std::vector<AddressPtr> &inputs, const st
auto args_size = static_cast<uint32_t>(UlongToUint(sizeof(void *)) * runtimeargs.size());
auto node_info = cnode->fullname_with_scope();
if (kernel_pack_->kernel_json_info().has_kernel_list) {
const auto dev_func = std::to_string(tiling_key_);
const auto kernel_info = node_info + "/" + std::to_string(tiling_key_);
// cppcheck-suppress unreadVariable
auto lock = device::KernelRuntime::LockRuntime(stream_ptr);
rtArgsEx_t args_info = {};
args_info.args = runtimeargs.data();
args_info.argsSize = args_size;
auto ret = rtKernelLaunchWithHandleV2(handle_, dev_func.c_str(), block_dim_, &args_info, l2ctrl, stream_ptr,
kernel_info.c_str());
auto ret =
rtKernelLaunchWithHandle(handle_, tiling_key_, block_dim_, &args_info, l2ctrl, stream_ptr, kernel_info.c_str());
if (ret != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Call runtime rtKernelLaunchWithHandleV2 error. Node info: " << node_info;
MS_LOG(ERROR) << "Call runtime rtKernelLaunchWithHandle error. Node info: " << node_info;
return false;
}
} else {

1
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/tiling/op_tiling_adapter.cc
View File

@ -175,7 +175,6 @@ void OpTilingCalculateAdapter::ConvertAtomicCompileInfo(const CNodePtr &node, ::
(void)::ge::AttrUtils::SetStr(*(*op_desc), ATOMIC_COMPILE_INFO_KEY, atomic_info_key);
(void)::ge::AttrUtils::SetStr(*(*op_desc), ATOMIC_COMPILE_INFO_JSON, atomic_compile_info);
}
// clean output
if (has_output) {
vector<int64_t> output_indexs;

3
mindspore/ccsrc/plugin/device/ascend/optimizer/buffer_fusion/reduce_eltwise_fusion_pass.cc
View File

@ -81,7 +81,8 @@ void ReduceEltwiseFusionPass::MatchSingleFusionPattern(const session::KernelGrap
auto cnode = node->cast<CNodePtr>();
MS_EXCEPTION_IF_NULL(cnode);
if (AnfAlgo::GetKernelType(cnode) == KernelType::TBE_KERNEL &&
AnfAlgo::GetFusionType(cnode) == kernel::FusionType::ELEMWISE && cnode->inputs().size() == ELTWISE_INPUT_SIZE) {
AnfAlgo::GetFusionType(cnode) == kernel::FusionType::ELEMWISE && cnode->inputs().size() == ELTWISE_INPUT_SIZE &&
common::AnfAlgo::GetCNodeName(cnode) != kCastOpName) {
MatchReduceEltwise(cnode, kernel_graph, candidate_fusion);
}
}

1
mindspore/lite/CMakeLists.txt
View File

@ -479,6 +479,7 @@ if(MSLITE_ENABLE_ACL)
set(ASCEND_RUNTIME_PATH ${ASCEND_PATH}/fwkacllib/lib64)
set(ASCEND_TOOLKIT_RUNTIME_PATH ${ASCEND_PATH}/ascend-toolkit/latest/fwkacllib/lib64)
set(ASCEND_CANN_RUNTIME_PATH ${ASCEND_PATH}/latest/fwkacllib/lib64)
set(ASCEND_CANN_RUNTIME_NEW_PATH ${ASCEND_PATH}/latest/lib64)
endif()
set(PKG_NAME_PREFIX mindspore-lite-${VERSION_STR})

3
mindspore/lite/src/CMakeLists.txt
View File

@ -422,7 +422,8 @@ add_dependencies(lite_src_mid fbs_src)
if(MSLITE_ENABLE_ACL AND NOT MSLITE_ENABLE_CLOUD_FUSION_INFERENCE)
include_directories(${TOP_DIR}/graphengine/inc/external)
add_subdirectory(extendrt/kernel/ascend)
link_directories(${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
link_directories(${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
endif()
add_library(mindspore-lite SHARED $<TARGET_OBJECTS:lite_src_mid>)

3
mindspore/lite/src/extendrt/kernel/ascend/CMakeLists.txt
View File

@ -1,6 +1,7 @@
include_directories(${TOP_DIR}/graphengine/inc/external)
find_library(ge_graph libgraph.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
find_library(ge_graph libgraph.so ${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
aux_source_directory(src ACL_SRC)
add_library(ascend_kernel_mid OBJECT ${ACL_SRC})

3
mindspore/lite/tools/converter/CMakeLists.txt
View File

@ -103,7 +103,8 @@ if(MSLITE_ENABLE_ACL)
include(${TOP_DIR}/cmake/dependency_graphengine.cmake)
endif()
add_subdirectory(adapter/acl)
link_directories(${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH})
link_directories(${ASCEND_RUNTIME_PATH} ${ASCEND_TOOLKIT_RUNTIME_PATH} ${ASCEND_CANN_RUNTIME_PATH}
${ASCEND_CANN_RUNTIME_NEW_PATH})
endif()
file(GLOB CXX_API_SRCS

29
mindspore/python/mindspore/_extends/parallel_compile/tbe_compiler/tbe_adapter.py
View File

@ -21,7 +21,7 @@ import traceback
from datetime import datetime
from tbe.common.rl_bank.bank_manager import set_current_op_name
from tbe.common.repository_manager.interface import cann_kb_unload, cann_kb_load
from tbe.common.repository_manager.interface import cann_kb_finalize, cann_kb_init
from tbe.common.rl_bank.bank_cfg import LocalLock
from te.platform.cce_conf import te_set_version
from te.platform.cce_policy import set_L1_info
@ -74,32 +74,25 @@ def _tune_init(job: TbeJob):
return res
def _cann_kb_load(job: TbeJob):
def _cann_kb_init(job: TbeJob):
"""
database load
database init
:param job:
:return:
"""
soc_version = job.soc_version
core_num = job.core_num
op_bank_path = job.op_bank_path
sys_config = {"soc_version": job.soc_version, "core_num": job.core_num}
load_config = {"op_bank_path": job.op_bank_path}
kb_type = None
res = cann_kb_load(soc_version, core_num, op_bank_path, kb_type)
res = cann_kb_init(sys_config, load_config, kb_type)
return res
def _cann_kb_unload(job: TbeJob):
def _cann_kb_finalize():
"""
database unload
:param job:
database finalize
:return:
"""
if job is None:
return 0
soc_version = job.soc_version
core_num = job.core_num
kb_type = None
res = cann_kb_unload(soc_version, core_num, kb_type)
res = cann_kb_finalize()
return res
@ -258,7 +251,7 @@ def tbe_initialize(job: TbeJob):
local_lock = LocalLock(lock_file)
try:
local_lock.lock()
res = _cann_kb_load(job)
res = _cann_kb_init(job)
if res == 1:
job.error("Cann kb load failed")
res = _parallel_compilation_init(job)
@ -668,7 +661,7 @@ def tbe_finalize(auto_tiling_mode, offline_tune, job: TbeJob):
if "RL" in auto_tiling_mode or offline_tune:
from schedule_search.rl_online_tune import rl_tune_deinit
rl_tune_deinit()
res = _cann_kb_unload(job)
res = _cann_kb_finalize()
if res == 1:
job.error("Cann kb unload failed")
return False

1
mindspore/python/mindspore/_extends/parallel_compile/tbe_compiler/tbe_helper.py
View File

@ -214,6 +214,7 @@ def get_options_info(job_content):
options["deviceId"] = job_content["SocInfo"]["deviceId"]
options["autoTilingMode"] = job_content["SocInfo"]["autoTilingMode"]
options["op_impl_mode_list"] = job_content["SocInfo"]["op_impl_mode_list"]
options["status_check"] = "false"
return options

5
mindspore/python/mindspore/ops/_op_impl/tbe/gru_v2_hidden_grad_cell.py
View File

@ -33,15 +33,16 @@ gru_v2_hidden_grad_cell_op_info = TBERegOp("GRUV2HiddenGradCell") \
.input(5, "reset", False, "required", "all") \
.input(6, "new", False, "required", "all") \
.input(7, "hidden_new", False, "required", "all") \
.input(8, "seq_mask", False, "optional", "all") \
.output(0, "dh_preh", False, "required", "all") \
.output(1, "dgate_h", False, "required", "all") \
.output(2, "dnt_x", False, "required", "all") \
.dtype_format(DataType.F32_FracNZ, DataType.F32_FracNZ, DataType.F32_FracNZ, DataType.F32_FracNZ,
DataType.F32_FracNZ, DataType.F32_FracNZ, DataType.F32_FracNZ, DataType.F32_FracNZ,
DataType.F32_FracNZ, DataType.F32_FracNZ, DataType.F32_FracNZ) \
DataType.F32_FracNZ, DataType.F32_FracNZ, DataType.F32_FracNZ, DataType.F16_FracNZ) \
.dtype_format(DataType.F16_FracNZ, DataType.F16_FracNZ, DataType.F16_FracNZ, DataType.F16_FracNZ,
DataType.F16_FracNZ, DataType.F16_FracNZ, DataType.F16_FracNZ, DataType.F16_FracNZ,
DataType.F16_FracNZ, DataType.F16_FracNZ, DataType.F16_FracNZ) \
DataType.F16_FracNZ, DataType.F16_FracNZ, DataType.F16_FracNZ, DataType.F16_FracNZ) \
.get_op_info()

5
mindspore/python/mindspore/profiler/parser/flops_parser.py
View File

@ -124,6 +124,8 @@ class FlopsParser:
step_idx = 0
aicore_file_doc = os.path.join(self._input_dir, "data")
source_files = self._get_aicore_files(aicore_file_doc)
if not source_files:
return
# parse all sliced aicore files.
for source_file in source_files:
source_file = validate_and_normalize_path(source_file)
@ -418,7 +420,8 @@ class FlopsParser:
"""Get aicore files."""
aicore_files = self._search_file(profiler_dir)
if not aicore_files:
raise ProfilerPathErrorException('The aicore file does not exist.')
logger.warning("Aicore file does not exist.")
return[]
return aicore_files

2
mindspore/python/mindspore/run_check/_check_version.py
View File

@ -215,7 +215,7 @@ class AscendEnvChecker(EnvChecker):
"""ascend environment check"""
def __init__(self):
self.version = ["1.81"]
self.version = ["1.82"]
atlas_nnae_version = "/usr/local/Ascend/nnae/latest/fwkacllib/version.info"
atlas_toolkit_version = "/usr/local/Ascend/ascend-toolkit/latest/fwkacllib/version.info"
hisi_fwk_version = "/usr/local/Ascend/latest/fwkacllib/version.info"

2
tests/st/networks/models/resnet50/test_resnet50_imagenet_and_thor.py
View File

@ -63,7 +63,7 @@ def thor_end():
thor_cost /= 4
print(f"resnet thor_loss: {thor_loss}, thor_cost: {thor_cost}")
assert thor_loss < 7
assert thor_cost < 30
assert thor_cost < 40
for i in range(4):
shutil.rmtree(os.path.join(sh_path, f"train_parallel{i+4}"))

20
tests/st/numpy_native/test_math_ops.py
View File

@ -469,7 +469,7 @@ def onp_average(x):
return a, b, c, d, e, f, g, h, i
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -1283,7 +1283,7 @@ def onp_kron(x, y):
return onp.kron(x, y)
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -1303,7 +1303,7 @@ def test_kron():
match_res(mnp.kron, onp.kron, x, y)
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -1437,7 +1437,7 @@ def onp_diff(input_array):
return a, b, c, d, e, f, g
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -1743,7 +1743,7 @@ def test_convolve():
match_all_arrays(mnp_res, onp_res)
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -2003,7 +2003,7 @@ def test_mean():
run_multi_test(mnp_mean, onp_mean, test_case.scalars, error=3)
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -2034,7 +2034,7 @@ def test_multi_dot():
match_all_arrays(mnp.multi_dot(mnp_arrays[1:-1]), onp.linalg.multi_dot(arrays[1:-1]))
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -2148,7 +2148,7 @@ def test_bincount():
onp.bincount(x, weights, minlength=25), error=3)
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -2472,7 +2472,7 @@ def test_result_type():
assert actual == expected
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@ -2555,7 +2555,7 @@ def test_ravel_multi_index():
match_array(actual.asnumpy(), expected, error=5)
@pytest.mark.level0
@pytest.mark.level1
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training

19
tests/ut/cpp/stub/profiling/profiling_stub.cc
View File

@ -90,3 +90,22 @@ ACL_FUNC_VISIBILITY aclprofConfig *aclprofCreateConfig(uint32_t *deviceIdList, u
}
ACL_FUNC_VISIBILITY aclError aclprofDestroyConfig(const aclprofConfig *profilerConfig) { return ACL_SUCCESS; }
/**
* @name profRegisterCallback
* @brief register callback to profiling
* @param moduleId [IN] module Id
* @param handle [IN] the pointer of callback
*/
MSVP_PROF_API int32_t MsprofRegisterCallback(uint32_t moduleId, ProfCommandHandle handle) { return 0; }
/*
* @name profReportData
* @brief start reporter/stop reporter/report date
* @param moduleId [IN] enum profReporterModuleId
* @param type [IN] enum profReporterCallbackType
* @param data [IN] data (nullptr on INTI/UNINIT)
* @param len [IN] data size (0 on INIT/UNINIT)
* @return enum MsprofErrorCod
*/
MSVP_PROF_API int32_t MsprofReportData(uint32_t moduleId, uint32_t type, void* data, uint32_t len) { return 0; }

19
tests/ut/cpp/stub/runtime/runtime_stub.cc
View File

@ -188,12 +188,6 @@ RTS_API rtError_t rtLabelCreateEx(rtLabel_t *label, rtStream_t stream) { return
RTS_API rtError_t rtLabelCreateExV2(rtLabel_t *lbl, rtModel_t mdl, rtStream_t stm) { return RT_ERROR_NONE; }
RTS_API rtError_t rtCpuKernelLaunchWithFlag(const void *soName, const void *kernelName, uint32_t blockDim,
const rtArgsEx_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stm,
uint32_t flags) {
return RT_ERROR_NONE;
}
RTS_API rtError_t rtCpuKernelLaunchWithFlagV2(const void *soName, const void *kernelName, uint32_t blockDim,
const rtArgsEx_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stm,
uint32_t flags) {
return RT_ERROR_NONE;
@ -210,11 +204,6 @@ RTS_API rtError_t rtProfilerTraceEx(uint64_t id, uint64_t modelId, uint16_t tagI
}
RTS_API rtError_t rtKernelLaunchWithFlag(const void *stubFunc, uint32_t blockDim, rtArgsEx_t *argsInfo,
rtSmDesc_t *smDesc, rtStream_t stm, uint32_t flags) {
return RT_ERROR_NONE;
}
RTS_API rtError_t rtKernelLaunchWithFlagV2(const void *stubFunc, uint32_t blockDim, rtArgsEx_t *argsInfo,
rtSmDesc_t *smDesc, rtStream_t stm, uint32_t flags) {
return RT_ERROR_NONE;
}
@ -225,14 +214,8 @@ RTS_API rtError_t rtProfRegisterCtrlCallback(uint32_t moduleId, rtProfCtrlHandle
RTS_API rtError_t rtGetRtCapability(rtFeatureType_t, int32_t, int64_t *) { return RT_ERROR_NONE; }
RTS_API rtError_t rtKernelLaunchWithHandle(void *hdl, const void *kernelInfoExt, uint32_t blockDim,
RTS_API rtError_t rtKernelLaunchWithHandle(void *hdl, const uint64_t tilingKey, uint32_t blockDim,
rtArgsEx_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stm,
const void *kernelInfo) {
return RT_ERROR_NONE;
}
RTS_API rtError_t rtKernelLaunchWithHandleV2(void *hdl, const void *kernelInfoExt, uint32_t blockDim,
rtArgsEx_t *argsInfo, rtSmDesc_t *smDesc, rtStream_t stm,
const void *kernelInfo) {
return RT_ERROR_NONE;
}

2
version.txt
View File

@ -1 +1 @@
1.7.0
1.8.0