m53297601
/
mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

!3489 use kernelruntime::mem_manager to reduce rtMalloc and rtFree time in trans data format

Browse Source 
Merge pull request !3489 from lvchangquan/master
pull/3489/MERGE
mindspore-ci-bot 5 years ago committed by Gitee
parent fcdad59ce6 fdbe4c19ba
commit d4b52ac59f
3 changed files with 37 additions and 33 deletions
Show Stats Download Patch File Download Diff File

48
mindspore/ccsrc/runtime/device/ascend/ascend_device_address.cc
Unescape View File

@ -153,6 +153,16 @@ bool SyncDeviceToHostAndFloatToFloat64(void *dst, size_t dst_size, const void *s
return true;
}
DeviceAddressPtr AssignLaunchMemory(size_t size, const std::string &format, TypeId type) {
auto ms_context = MsContext::GetInstance();
MS_EXCEPTION_IF_NULL(ms_context);
auto device_id = ms_context->device_id();
auto runtime_instance = device::KernelRuntimeManager::Instance().GetKernelRuntime(kAscendDevice, device_id);
MS_EXCEPTION_IF_NULL(runtime_instance);
auto address_ptr = runtime_instance->AssignSingleOpLaunchMemory(size, format, type);
return address_ptr;
}
size_t GetCommonAlignSize(size_t input_size) {
return (input_size + kMemAlignSize + 31) / kMemAlignSize * kMemAlignSize;
}
@ -325,18 +335,15 @@ void AscendDeviceAddress::LaunchTransData(kernel::KernelModPtr kernel_mod_ptr, v
AddressPtrList kernel_inputs = {input_address};
AddressPtrList kernel_outputs = {output_address};
AddressPtrList kernel_workspaces;
std::vector<void *> workspaces_address_ptr(workspace_size_list.size(), nullptr);
if (!workspace_size_list.empty()) {
for (size_t i = 0; i < workspace_size_list.size(); ++i) {
auto workspace_size = GetCommonAlignSize(workspace_size_list[i]);
auto ret_malloc = rtMalloc(&workspaces_address_ptr[i], workspace_size, RT_MEMORY_HBM);
if (ret_malloc != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Failed to rtMalloc memory";
}
auto workspace_address_ptr = AssignLaunchMemory(workspace_size, "", kTypeUnknown);
MS_EXCEPTION_IF_NULL(workspace_address_ptr);
auto workspace_address = std::make_shared<kernel::Address>();
MS_EXCEPTION_IF_NULL(workspace_address);
workspace_address->addr = workspaces_address_ptr[i];
workspace_address->size = workspace_size;
workspace_address->addr = workspace_address_ptr->GetMutablePtr();
workspace_address->size = workspace_address_ptr->GetSize();
kernel_workspaces.push_back(workspace_address);
}
}
@ -350,15 +357,6 @@ void AscendDeviceAddress::LaunchTransData(kernel::KernelModPtr kernel_mod_ptr, v
if (!ret) {
MS_LOG(ERROR) << "Launch kernel failed.";
}
SyncStream();
if (!workspace_size_list.empty()) {
for (size_t i = 0; i < workspace_size_list.size(); ++i) {
auto ret_free = rtFree(workspaces_address_ptr[i]);
if (ret_free != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Failed to rtFree memory";
}
}
}
}
kernel::KernelModPtr AscendDeviceAddress::CompileTransDataAndObtainKernelMod(const nlohmann::json &kernel_json) const {
@ -418,19 +416,17 @@ bool AscendDeviceAddress::SyncDeviceToHostAndConvertFormatBasedOnTransData(const
size = device_dtype_size * shape_size;
}
size = GetCommonAlignSize(size);
void *output_address_ptr = nullptr;
auto ret_malloc = rtMalloc(&output_address_ptr, size, RT_MEMORY_HBM);
if (ret_malloc != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Failed to rtMalloc memory";
}
auto output_address = AssignLaunchMemory(size, kOpFormat_NCHW, type_id_);
MS_EXCEPTION_IF_NULL(output_address);
auto workspace_size_list = GetWorkspaceSizeList(kernel_json);
// launch
LaunchTransData(kernel_mod_ptr, output_address_ptr, size, workspace_size_list);
LaunchTransData(kernel_mod_ptr, output_address->GetMutablePtr(), output_address->GetSize(), workspace_size_list);
SyncStream();
if (type_id_ == type) {
SyncMemory(host_ptr, output_address_ptr, host_size, RT_MEMCPY_DEVICE_TO_HOST);
SyncMemory(host_ptr, output_address->GetPtr(), host_size, RT_MEMCPY_DEVICE_TO_HOST);
} else {
auto host = std::vector<uint8_t>(size);
SyncMemory(host.data(), output_address_ptr, size, RT_MEMCPY_DEVICE_TO_HOST);
SyncMemory(host.data(), output_address->GetPtr(), size, RT_MEMCPY_DEVICE_TO_HOST);
auto shape_size = trans::ShapeSize(host_shape);
const trans::TypeIdArgs type_args{host.data(), shape_size, type_id_, type, host_size};
sync_ok = trans::TransDataType(type_args, host_ptr);
@ -439,10 +435,6 @@ bool AscendDeviceAddress::SyncDeviceToHostAndConvertFormatBasedOnTransData(const
return false;
}
}
auto ret_free = rtFree(output_address_ptr);
if (ret_free != RT_ERROR_NONE) {
MS_LOG(ERROR) << "Failed to rtFree memory";
}
return sync_ok;
}

16
mindspore/ccsrc/runtime/device/kernel_runtime.cc
Unescape View File

@ -842,9 +842,10 @@ void KernelRuntime::ClearGraphRuntimeResource(uint32_t graph_id) {
MS_LOG(INFO) << "Clear graph:" << graph_id << " runtime resource";
}
bool KernelRuntime::LaunchTaskBasedOnSingleKernel(kernel::KernelModPtr kernel_mod_ptr, AddressPtrList kernel_inputs,
AddressPtrList kernel_outputs,
AddressPtrList kernel_workspaces) const {
bool KernelRuntime::LaunchTaskBasedOnSingleKernel(kernel::KernelModPtr kernel_mod_ptr,
const AddressPtrList &kernel_inputs,
const AddressPtrList &kernel_outputs,
const AddressPtrList &kernel_workspaces) const {
MS_EXCEPTION_IF_NULL(kernel_mod_ptr);
auto ret = kernel_mod_ptr->Launch(kernel_inputs, kernel_workspaces, kernel_outputs, stream_);
if (!ret) {
@ -854,6 +855,15 @@ bool KernelRuntime::LaunchTaskBasedOnSingleKernel(kernel::KernelModPtr kernel_mo
return true;
}
DeviceAddressPtr KernelRuntime::AssignSingleOpLaunchMemory(size_t size, const std::string &format, TypeId type) {
auto device_address = CreateDeviceAddress(nullptr, size, format, type);
MS_EXCEPTION_IF_NULL(device_address);
MS_EXCEPTION_IF_NULL(mem_manager_);
auto base_ptr = mem_manager_->MallocMem(kDynamicMem, size);
device_address->set_ptr(base_ptr);
return device_address;
}
#ifdef ENABLE_DUMP_E2E
bool KernelRuntime::SetDumpConf() {
dump_conf_ptr_ = std::make_shared<Dump>();

6
mindspore/ccsrc/runtime/device/kernel_runtime.h
Unescape View File

@ -65,8 +65,9 @@ class KernelRuntime {
virtual bool RunTask(const session::KernelGraph *graph);
virtual bool GenTask(const session::KernelGraph *graph);
bool LaunchKernel(const session::KernelGraph *graph);
bool LaunchTaskBasedOnSingleKernel(kernel::KernelModPtr kernel_mod_ptr, AddressPtrList kernel_inputs,
AddressPtrList kernel_outputs, AddressPtrList kernel_workspaces) const;
bool LaunchTaskBasedOnSingleKernel(kernel::KernelModPtr kernel_mod_ptr, const AddressPtrList &kernel_inputs,
const AddressPtrList &kernel_outputs,
const AddressPtrList &kernel_workspaces) const;
virtual void AssignStaticMemoryInput(const session::KernelGraph *graph);
virtual void AssignStaticMemoryValueNode(session::KernelGraph *graph);
virtual void ClearGraphRuntimeResource(uint32_t graph_id);
@ -79,6 +80,7 @@ class KernelRuntime {
// for GPU and D to impl
virtual void ReleaseDeviceRes() {}
void set_device_id(uint32_t device_id) { device_id_ = device_id; }
DeviceAddressPtr AssignSingleOpLaunchMemory(size_t size, const std::string &format, TypeId type);
protected:
virtual DeviceAddressPtr CreateDeviceAddress(void *device_ptr, size_t device_size, const string &format,

Write Preview
Loading…
Cancel
Save