m53297601
/
graphengine
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '957b435f16'
${ noResults }
graphengine/ge/graph/load/new_model_manager/davinci_model.cc

4202 lines
168 KiB
Raw Blame History

/**
* Copyright 2019-2020 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 "graph/load/new_model_manager/davinci_model.h"
#include <cce/dnn.h>
#include <dlfcn.h>
#include <graph/utils/node_utils.h>
#include <pthread.h>
#include <sched.h>
#include <sys/prctl.h>
#include <algorithm>
#include <map>
#include <utility>
#include "common/debug/log.h"
#include "common/formats/formats.h"
#include "common/formats/utils/formats_trans_utils.h"
#include "common/math/math_util.h"
#include "common/op/ge_op_utils.h"
#include "common/profiling/profiling_manager.h"
#include "common/properties_manager.h"
#include "common/scope_guard.h"
#include "common/thread_pool.h"
#include "framework/common/debug/ge_log.h"
#include "graph/common/ge_call_wrapper.h"
#include "graph/compute_graph.h"
#include "graph/debug/ge_attr_define.h"
#include "graph/ge_context.h"
#include "graph/graph.h"
#include "graph/load/new_model_manager/cpu_queue_schedule.h"
#include "graph/load/new_model_manager/model_manager.h"
#include "graph/load/new_model_manager/tbe_handle_store.h"
#include "graph/manager/graph_mem_allocator.h"
#include "graph/manager/graph_var_manager.h"
#include "graph/manager/trans_var_data_utils.h"
#include "graph/manager/util/debug.h"
#include "graph/model_serialize.h"
#include "graph/node.h"
#include "graph/utils/graph_utils.h"
#include "graph/utils/type_utils.h"
#include "init/gelib.h"
#include "mmpa/mmpa_api.h"
#include "omm/csa_interact.h"
#include "runtime/base.h"
#include "runtime/dev.h"
#include "runtime/event.h"
#include "runtime/mem.h"
#include "runtime/rt_model.h"
#include "runtime/stream.h"
#include "securec.h"
#include "graph/common/local_context.h"
#include "common/formats/utils/formats_trans_utils.h"
// create std::thread, catch exceptions using try/catch
#define CREATE_STD_THREAD(thread_id, func, args) \
do { \
try { \
thread_id = std::thread(func, args); \
} catch (const std::system_error &e) { \
GELOGE(FAILED, "Caught system_error with code:%d, meaning:%s", e.code().value(), e.what()); \
GELOGE(FAILED, "Thread creat FAIL, Please check the left resource!"); \
return FAILED; \
} \
} while (0)
namespace ge {
namespace {
const uint32_t kDataIndex = 0;
const uint32_t kOutputNum = 1;
const uint32_t kTrueBranchStreamNum = 1;
const uint32_t kGetDynamicDimsCount = 1;
const uint32_t kThreadNum = 16;
const uint32_t kAddrLen = sizeof(void *);
const int kDecimal = 10;
const int kBytes = 8;
const uint32_t kDataMemAlignSizeCompare = 64;
const uint32_t kDumpL1FusionOpMByteSize = 2 * 1024 * 1024;
const uint32_t kDumpFlagOfL1Fusion = 0;
const char *const kDefaultBatchLable = "Batch_default";
const char *const kGetDynamicDimsName = "ascend_mbatch_get_dynamic_dims_node";
const int32_t kInvalidStream = -1;
const uint32_t kEndOfSequence = 0x0704000a;
const uint32_t kEndOfSequenceNew = 507005;
inline bool IsDataOp(const std::string &node_type) {
return node_type == DATA_TYPE || node_type == AIPP_DATA_TYPE || node_type == ANN_DATA_TYPE;
}
inline bool IsNoTaskAndDumpNeeded(const OpDescPtr &op_desc) {
bool save_dump_info = false;
(void)ge::AttrUtils::GetBool(op_desc, ATTR_NO_TASK_AND_DUMP_NEEDED, save_dump_info);
return save_dump_info;
}
} // namespace
std::mutex DavinciModel::tvm_bin_mutex_;
DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener> &listener)
: weights_mem_base_(nullptr),
var_mem_base_(nullptr),
mem_base_(nullptr),
is_inner_mem_base_(false),
is_inner_weight_base_(false),
is_inner_p2p_mem_base_(false),
data_inputer_(nullptr),
load_begin_time_(0),
load_end_time_(0),
time_info_(),
dataInputTid(0),
is_model_has_inited_(false),
model_id_(0),
runtime_model_id_(0),
version_(0),
ge_model_(nullptr),
thread_id_(),
listener_(listener),
run_flg_(false),
priority_(priority),
rt_model_handle_(nullptr),
rt_model_stream_(nullptr),
is_inner_model_stream_(false),
is_async_mode_(false),
last_execute_mode_(INITIALIZATION),
session_id_(0),
device_id_(0),
maxDumpOpNum_(0), data_dumper_(runtime_param_),
iterator_count_(0),
is_l1_fusion_enable_(false),
is_first_execute_(true) {
op_list_.clear();
}
DavinciModel::~DavinciModel() {
try {
Status ret = data_dumper_.UnloadDumpInfo();
if (ret != SUCCESS) {
GELOGW("UnloadDumpInfo failed, ret: %u.", ret);
}
for (const auto &op_and_addr : saved_task_addrs_) {
auto addr = op_and_addr.second;
if (addr != nullptr) {
GE_CHK_RT(rtFree(addr));
}
addr = nullptr;
}
saved_task_addrs_.clear();
GE_CHK_STATUS(ModelRunStop());
op_list_.clear();
data_op_list_.clear();
output_op_list_.clear();
tensor_name_to_fixed_addr_size_.clear();
tensor_name_to_peer_output_index_.clear();
GE_DELETE_NEW_SINGLE(data_inputer_);
// check rt ctx is exist. rt api call will cause error log when ctx not exist
rtContext_t ctx = nullptr;
rtError_t rt_ret = rtCtxGetCurrent(&ctx);
if (rt_ret == RT_ERROR_NONE) {
UnbindTaskSinkStream();
for (size_t i = 0; i < label_list_.size(); ++i) {
if (label_list_[i] != nullptr) {
GE_LOGW_IF(rtLabelDestroy(label_list_[i]) != RT_ERROR_NONE, "Destroy label failed, index: %zu", i);
}
}
for (size_t i = 0; i < stream_list_.size(); ++i) {
GE_LOGW_IF(rtStreamDestroy(stream_list_[i]) != RT_ERROR_NONE, "Destroy stream failed, index: %zu", i);
}
for (size_t i = 0; i < event_list_.size(); ++i) {
GE_LOGW_IF(rtEventDestroy(event_list_[i]) != RT_ERROR_NONE, "Destroy event failed, index: %zu", i);
}
FreeWeightsMem();
FreeFeatureMapMem();
FreeP2PMem();
if (l1_fusion_addr_ != nullptr) {
GE_CHK_RT(rtFree(l1_fusion_addr_));
}
if (rt_model_handle_ != nullptr) {
GE_CHK_RT(rtModelDestroy(rt_model_handle_));
rt_model_handle_ = nullptr;
}
}
OpDebugUnRegister();
GELOGI("do ReleaseTask");
ReleaseTask();
CleanTbeHandle();
var_mem_base_ = nullptr;
if (known_node_) {
if (args_ != nullptr) {
GE_CHK_RT(rtFree(args_));
}
total_io_addrs_.clear();
if (fixed_addrs_ != nullptr) {
GE_CHK_RT(rtFree(fixed_addrs_));
}
}
} catch (...) {
GELOGW("DavinciModel::~DavinciModel: clear op_list catch exception.");
}
}
void DavinciModel::UnbindHcomStream() {
if (!all_hccl_stream_list_.empty()) {
for (size_t i = 0; i < all_hccl_stream_list_.size(); i++) {
GE_LOGW_IF(rtModelUnbindStream(rt_model_handle_, all_hccl_stream_list_[i]) != RT_ERROR_NONE,
"Unbind hccl stream from model failed! Index: %zu", i);
GE_LOGW_IF(rtStreamDestroy(all_hccl_stream_list_[i]) != RT_ERROR_NONE, "Destroy hccl stream for rt_model failed!")
}
}
return;
}
void DavinciModel::ReleaseTask() {
for (const auto &task : cpu_task_list_) {
if (task != nullptr) {
GE_CHK_STATUS(task->Release(), "Release task failed.");
}
}
cpu_task_list_.clear();
for (const auto &task : task_list_) {
if (task != nullptr) {
GE_CHK_STATUS(task->Release(), "Release task failed.");
}
}
}
Status DavinciModel::Assign(const GeModelPtr &ge_model) {
if (ge_model == nullptr) {
GELOGI("can't assign null ge_model");
return FAILED;
}
ge_model_ = ge_model;
return SUCCESS;
}
///
/// @ingroup ge
/// @brief Reduce memory usage after task sink.
/// @return: void
///
void DavinciModel::Shrink() {
ge_model_.reset(); // delete object.
}
Status DavinciModel::InitModelMem(void *dev_ptr, size_t mem_size, void *weight_ptr, size_t weight_size) {
if (is_model_has_inited_) {
GELOGE(FAILED, "call InitModelMem more than once .");
return FAILED;
}
is_model_has_inited_ = true;
std::size_t data_size = TotalMemSize();
std::size_t p2p_data_size = P2PMemInfos().at(RT_MEMORY_P2P_DDR).memory_size;
const Buffer &weights = ge_model_->GetWeight();
std::size_t weights_size = weights.GetSize();
GE_CHECK_LE(weights_size, ALLOC_MEMORY_MAX_SIZE);
if ((dev_ptr != nullptr) && (mem_size < TotalMemSize())) {
GELOGE(FAILED, "Invalid mem param: mem_size=%zu totalsize=%zu.", mem_size, TotalMemSize());
return FAILED;
}
if ((weight_ptr != nullptr) && (weight_size < weights_size)) {
GELOGE(FAILED, "Invalid mem param: weight_size=%zu totalsize=%zu.", weight_size, weights_size);
return FAILED;
}
mem_base_ = static_cast<uint8_t *>(dev_ptr);
p2p_mem_base_ = static_cast<uint8_t *>(dev_ptr);
weights_mem_base_ = static_cast<uint8_t *>(dev_ptr);
is_inner_mem_base_ = false;
is_inner_weight_base_ = false;
if (TotalMemSize() && mem_base_ == nullptr) {
mem_base_ = MallocFeatureMapMem(data_size);
if (mem_base_ == nullptr) {
GELOGE(GE_EXEC_ALLOC_FEATURE_MAP_MEM_FAILED, "Alloc feature map memory failed. size: %zu", data_size);
return GE_EXEC_ALLOC_FEATURE_MAP_MEM_FAILED;
}
GEEVENT("[IMAS]InitModelMem graph_%u MallocMemory type[F] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id,
mem_base_, data_size);
weights_mem_base_ = mem_base_;
is_inner_mem_base_ = true;
is_inner_weight_base_ = true;
}
if (p2p_data_size != 0) {
p2p_mem_base_ = MallocP2PMem(p2p_data_size);
if (p2p_mem_base_ == nullptr) {
GELOGE(GE_EXEC_ALLOC_P2P_MEM_FAILED, "Alloc p2p memory failed,size: %zu", p2p_data_size);
return GE_EXEC_ALLOC_P2P_MEM_FAILED;
}
GELOGI("InitModelMem graph_%u MallocMemory type[P] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id,
p2p_mem_base_, p2p_data_size);
is_inner_p2p_mem_base_ = true;
}
if (weights_size != 0) {
weights_mem_base_ = static_cast<uint8_t *>(weight_ptr);
is_inner_weight_base_ = false;
if (weight_ptr == nullptr) {
weights_mem_base_ = MallocWeightsMem(weights_size);
if (weights_mem_base_ == nullptr) {
GELOGE(GE_EXEC_ALLOC_WEIGHT_MEM_FAILED, "Alloc weight memory failed. size: %zu", weights_size);
return GE_EXEC_ALLOC_WEIGHT_MEM_FAILED;
}
is_inner_weight_base_ = true;
}
GELOGI("[IMAS]InitModelMem graph_%u MallocMemory type[W] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id,
weights_mem_base_, weights_size);
GE_CHK_RT_RET(rtMemcpy(weights_mem_base_, weights_size, weights.GetData(), weights_size, RT_MEMCPY_HOST_TO_DEVICE));
GELOGI("copy weights data to device");
}
GE_CHK_STATUS_RET(InitVariableMem(), "Init variable memory failed.");
runtime_param_.mem_base = mem_base_;
runtime_param_.weight_base = weights_mem_base_;
runtime_param_.memory_infos[RT_MEMORY_P2P_DDR].memory_base = p2p_mem_base_;
return SUCCESS;
}
Status DavinciModel::InitVariableMem() {
// malloc variable memory base
var_mem_base_ = VarManager::Instance(session_id_)->GetVarMemoryBase(RT_MEMORY_HBM);
if (TotalVarMemSize() && var_mem_base_ == nullptr) {
Status ret = VarManager::Instance(session_id_)->MallocVarMemory(TotalVarMemSize());
if (ret != SUCCESS) {
GELOGE(ret, "Malloc variable memory failed.");
return ret;
}
var_mem_base_ = VarManager::Instance(session_id_)->GetVarMemoryBase(RT_MEMORY_HBM);
GEEVENT("[IMAS]InitVariableMem graph_%u MallocMemory type[V] memaddr[%p] mem_size[%zu]", runtime_param_.graph_id,
var_mem_base_, TotalVarMemSize());
}
runtime_param_.var_base = var_mem_base_;
return SUCCESS;
}
void DavinciModel::InitRuntimeParams() {
int64_t value = 0;
bool ret;
MemInfo p2p_mem_info;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_MEMORY_SIZE, value);
runtime_param_.mem_size = ret ? (uint64_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_WEIGHT_SIZE, value);
runtime_param_.weight_size = ret ? (uint64_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_STREAM_NUM, value);
runtime_param_.stream_num = ret ? (uint32_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_EVENT_NUM, value);
runtime_param_.event_num = ret ? (uint32_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_LABEL_NUM, value);
runtime_param_.label_num = ret ? (uint32_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_BATCH_NUM, value);
runtime_param_.batch_num = ret ? (uint32_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, MODEL_ATTR_TASK_GEN_BASE_ADDR, value);
runtime_param_.logic_mem_base = ret ? (uint64_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, MODEL_ATTR_TASK_GEN_WEIGHT_ADDR, value);
runtime_param_.logic_weight_base = ret ? (uint64_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ge::MODEL_ATTR_SESSION_ID, value);
runtime_param_.session_id = ret ? (uint64_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_TASK_GEN_VAR_ADDR, value);
runtime_param_.logic_var_base = ret ? (uint64_t)value : 0;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_VAR_SIZE, value);
runtime_param_.var_size = ret ? (uint64_t)value : 0;
session_id_ = runtime_param_.session_id;
ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_P2P_MEMORY_SIZE, value);
p2p_mem_info.memory_size = ret ? (uint64_t)value : 0;
runtime_param_.memory_infos[RT_MEMORY_P2P_DDR] = std::move(p2p_mem_info);
GELOGI(
"InitRuntimeParams(), session_id:%lu, stream_num:%u, event_num:%u, label_num:%u, "
"logic_mem_base:0x%lx, logic_weight_base:0x%lx, logic_var_base:0x%lx, "
"memory_size:%lu, weight_size:%lu, var_size:%lu",
runtime_param_.session_id, runtime_param_.stream_num, runtime_param_.event_num, runtime_param_.label_num,
runtime_param_.logic_mem_base, runtime_param_.logic_weight_base, runtime_param_.logic_var_base,
runtime_param_.mem_size, runtime_param_.weight_size, runtime_param_.var_size);
}
void DavinciModel::CheckHasHcomOp() {
Graph graph = ge_model_->GetGraph();
auto compute_graph = GraphUtils::GetComputeGraph(graph);
if (compute_graph == nullptr) {
return;
}
for (const auto &node : compute_graph->GetAllNodes()) {
OpDescPtr op_desc = node->GetOpDesc();
GE_IF_BOOL_EXEC(op_desc == nullptr, GELOGW("Node OpDesc is nullptr"); continue);
GE_IF_BOOL_EXEC(((op_desc->GetType() == HCOMBROADCAST) || (op_desc->GetType() == HCOMALLGATHER) ||
(op_desc->GetType() == HCOMALLREDUCE) || (op_desc->GetType() == HCOMSEND) ||
(op_desc->GetType() == HCOMRECEIVE) || (op_desc->GetType() == HCOMREDUCESCATTER) ||
(op_desc->GetType() == HVDCALLBACKALLREDUCE) || (op_desc->GetType() == HVDCALLBACKALLGATHER) ||
(op_desc->GetType() == HVDCALLBACKBROADCAST) || (op_desc->GetType() == HVDWAIT)),
uint32_t stream_id = static_cast<uint32_t>(op_desc->GetStreamId());
(void)hcom_streams_.emplace(stream_id); GELOGD("hcom stream: %u.", stream_id); continue);
}
}
///
/// @ingroup ge
/// @brief Make active stream list and bind to model.
/// @return: 0 for success / others for fail
///
Status DavinciModel::BindModelStream() {
// Stream not in active_stream_indication_ is active stream.
is_stream_list_bind_ = false;
if ((!input_queue_ids_.empty() || !output_queue_ids_.empty()) || (deploy_type_ == AICPU_DEPLOY_CROSS_THREAD)) {
for (size_t i = 0; i < stream_list_.size(); ++i) {
if (active_stream_indication_.count(i) == 0) {
active_stream_list_.push_back(stream_list_[i]);
active_stream_indication_.insert(i); // deactive all model stream.
}
}
}
for (size_t i = 0; i < stream_list_.size(); ++i) {
if (active_stream_indication_.count(i) > 0) {
GELOGI("rtModelBindStream[%zu]", i);
GE_CHK_RT_RET(rtModelBindStream(rt_model_handle_, stream_list_[i], RT_INVALID_FLAG));
} else {
// bind rt_model_handel to all streams that relates to op
GE_CHK_RT_RET(rtModelBindStream(rt_model_handle_, stream_list_[i], RT_HEAD_STREAM));
}
}
is_stream_list_bind_ = true;
return SUCCESS;
}
Status DavinciModel::DoTaskSink() {
// task sink is supported as model_task_def is set
const auto &model_task_def = ge_model_->GetModelTaskDefPtr();
if (model_task_def == nullptr) {
return SUCCESS;
}
GE_CHK_RT_RET(rtGetAicpuDeploy(&deploy_type_));
GELOGI("do task_sink. AiCpu deploy type is: %x.", deploy_type_);
GE_CHK_STATUS_RET(BindModelStream(), "Bind model stream failed.");
if (known_node_) {
GE_CHK_STATUS_RET(MallocKnownArgs(), "Mallloc known node args failed.");
}
GE_CHK_STATUS_RET(InitTaskInfo(*model_task_def.get()), "InitTaskInfo failed.");
GE_CHK_STATUS_RET(ModelManager::GetInstance()->LaunchCustAicpuSo(), "Launch cust aicpu so failed.");
GE_CHK_STATUS_RET(InitEntryTask(), "InitEntryTask failed.");
GE_CHK_STATUS_RET(DistributeTask(), "Distribute failed.");
GE_CHK_RT_RET(rtModelLoadComplete(rt_model_handle_));
SetCopyOnlyOutput();
return SUCCESS;
}
// set device use aicore(0) or vectorcore(1)
Status DavinciModel::SetTSDevice() {
int64_t value = 0;
bool ret = ge::AttrUtils::GetInt(ge_model_, ATTR_MODEL_CORE_TYPE, value);
uint32_t core_type = ret ? static_cast<uint32_t>(value) : 0;
GELOGI("SetTSDevice: %u", core_type);
rtError_t rt_ret = rtSetTSDevice(core_type);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "SetTSDevice failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
return SUCCESS;
}
Status DavinciModel::OpDebugRegister() {
bool is_op_debug = false;
(void)ge::AttrUtils::GetBool(ge_model_, ATTR_OP_DEBUG_FLAG, is_op_debug);
GELOGD("The value of op_debug in ge_model_ is %d.", is_op_debug);
if (is_op_debug) {
debug_reg_mutex_.lock();
rtError_t rt_ret = rtMalloc(&op_debug_addr_, kOpDebugMemorySize, RT_MEMORY_DDR);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "rtMalloc error, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
uint64_t debug_addrs_tmp = static_cast<uint64_t>(reinterpret_cast<uintptr_t>(op_debug_addr_));
// For data dump, aicpu needs the pointer to pointer that save the real debug address.
rt_ret = rtMalloc(&p2p_debug_addr_, kDebugP2pSize, RT_MEMORY_HBM);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "rtMalloc error, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
rt_ret = rtMemcpy(p2p_debug_addr_, sizeof(uint64_t), &debug_addrs_tmp, sizeof(uint64_t), RT_MEMCPY_HOST_TO_DEVICE);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "rtMemcpy to p2p_addr error: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
uint32_t op_debug_mode = 0;
(void)ge::AttrUtils::GetInt(ge_model_, ATTR_OP_DEBUG_MODE, op_debug_mode);
GELOGD("The value of op_debug_mode in ge_model_ is %u.", op_debug_mode);
uint32_t debug_task_id = 0;
uint32_t debug_stream_id = 0;
rt_ret = rtDebugRegister(rt_model_handle_, op_debug_mode, op_debug_addr_, &debug_stream_id, &debug_task_id);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "rtDebugRegister error, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
GELOGI("debug_task_id:%d, debug_stream_id:%u", debug_task_id, debug_stream_id);
is_op_debug_reg_ = true;
data_dumper_.SaveOpDebugId(debug_task_id, debug_stream_id, p2p_debug_addr_, is_op_debug);
}
return SUCCESS;
}
void DavinciModel::OpDebugUnRegister() {
GELOGI("OpDebugUnRegister, is_op_debug_reg_ = %d", is_op_debug_reg_);
if (is_op_debug_reg_) {
debug_reg_mutex_.unlock();
rtError_t rt_ret = RT_ERROR_NONE;
if (rt_model_handle_ != nullptr) {
GELOGD("start call debug_unregister.");
rt_ret = rtDebugUnRegister(rt_model_handle_);
if (rt_ret != RT_ERROR_NONE) {
GELOGW("rtDebugUnRegister failed, ret: 0x%X", rt_ret);
}
}
if (op_debug_addr_ != nullptr) {
rt_ret = rtFree(op_debug_addr_);
if (rt_ret != RT_ERROR_NONE) {
GELOGW("rtFree failed, ret: 0x%X", rt_ret);
}
op_debug_addr_ = nullptr;
}
if (p2p_debug_addr_ != nullptr) {
rt_ret = rtFree(p2p_debug_addr_);
if (rt_ret != RT_ERROR_NONE) {
GELOGW("rtFree failed, ret: 0x%X", rt_ret);
}
p2p_debug_addr_ = nullptr;
}
is_op_debug_reg_ = false;
}
return;
}
// initialize op sequence and call initialization function of each op respectively
Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size_t weight_size) {
// validating params
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(priority_ < 0 || priority_ > 7, return PARAM_INVALID,
"Priority must between 0-7, now is %d", priority_);
GE_CHK_BOOL_RET_STATUS(ge_model_ != nullptr, PARAM_INVALID, "GeModel is null.");
Graph graph = ge_model_->GetGraph();
ComputeGraphPtr compute_graph = GraphUtils::GetComputeGraph(graph);
GE_CHK_BOOL_RET_STATUS(compute_graph != nullptr, INTERNAL_ERROR, "Get compute graph is nullptr.");
// Initializing runtime_param_
InitRuntimeParams();
// RTS set aicore or vectorcore
GE_CHK_STATUS_RET(SetTSDevice(), "SetTSDevice failed");
version_ = ge_model_->GetVersion();
name_ = ge_model_->GetName();
(void)ge::AttrUtils::GetBool(ge_model_, ATTR_NAME_SWITCH_FOR_L1_FUSION, is_l1_fusion_enable_);
GELOGD("The value of ge.l1Fusion in ge_model_ is %d.", is_l1_fusion_enable_);
CheckHasHcomOp();
vector<int64_t> huge_stream_list;
(void)ge::AttrUtils::GetListInt(ge_model_, ATTR_MODEL_HUGE_STREAM_LIST, huge_stream_list);
std::set<int64_t> huge_streams(huge_stream_list.begin(), huge_stream_list.end());
for (uint32_t i = 0; i < StreamNum(); i++) {
rtStream_t stream = nullptr;
GE_MAKE_GUARD_RTSTREAM(stream);
uint32_t stream_flags = RT_STREAM_PERSISTENT;
if (huge_streams.find(i) != huge_streams.end()) {
GELOGI("Stream %u is huge stream.", i);
stream_flags |= RT_STREAM_HUGE;
}
if (hcom_streams_.find(i) != hcom_streams_.end()) {
GE_CHK_RT_RET(rtStreamCreateWithFlags(&stream, priority_, stream_flags | RT_STREAM_FORCE_COPY));
} else {
GE_CHK_RT_RET(rtStreamCreateWithFlags(&stream, priority_, stream_flags));
}
GE_DISMISS_GUARD(stream);
stream_list_.push_back(stream);
int32_t rt_stream_id = kInvalidStream;
(void)rtGetStreamId(stream, &rt_stream_id);
GELOGI("Logical stream index:%u, stream:%p, rtstream: %d.", i, stream, rt_stream_id);
}
for (uint32_t i = 0; i < EventNum(); i++) {
rtEvent_t rt_event;
GE_CHK_RT_RET(rtEventCreate(&rt_event));
event_list_.push_back(rt_event);
}
label_list_.resize(LabelNum(), nullptr);
// create model_handle to load model
GE_CHK_RT_RET(rtModelCreate(&rt_model_handle_, 0));
GE_CHK_RT_RET(rtModelGetId(rt_model_handle_, &runtime_model_id_));
// inference will use default graph_id 0;
runtime_param_.graph_id = compute_graph->GetGraphID();
// op debug register
GE_CHK_STATUS_RET(OpDebugRegister(), "OpDebugRegister failed");
GE_TIMESTAMP_START(TransAllVarData);
GE_CHK_STATUS_RET(TransAllVarData(compute_graph, runtime_param_.graph_id), "TransAllVarData failed.");
GE_TIMESTAMP_END(TransAllVarData, "GraphLoader::TransAllVarData");
GE_CHK_STATUS_RET(TransVarDataUtils::CopyVarData(compute_graph, session_id_, device_id_), "copy var data failed.");
GE_TIMESTAMP_START(InitModelMem);
GELOGI("Known node is %d", known_node_);
if (!known_node_) {
GE_CHK_STATUS_RET_NOLOG(InitModelMem(dev_ptr, mem_size, weight_ptr, weight_size));
data_inputer_ = new (std::nothrow) DataInputer();
GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, MEMALLOC_FAILED, "data_inputer_ is nullptr.");
}
GE_TIMESTAMP_END(InitModelMem, "GraphLoader::InitModelMem");
for (const ge::NodePtr &node : compute_graph->GetDirectNode()) {
auto op_desc = node->GetOpDesc();
GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
GE_IF_BOOL_EXEC(op_desc->GetType() != VARIABLE, continue);
GE_IF_BOOL_EXEC(IsBroadCastOpData(node),
(void)ge::AttrUtils::SetStr(op_desc, VAR_ATTR_VAR_IS_BROADCAST, "var_is_restore"););
}
GE_CHK_STATUS_RET(InitNodes(compute_graph), "Init nodes failed");
SetDataDumperArgs(compute_graph);
GE_TIMESTAMP_START(DoTaskSink);
auto ret = DoTaskSink();
GE_TIMESTAMP_END(DoTaskSink, "GraphLoader::DoTaskSink");
auto all_dump_model = GetDumpProperties().GetAllDumpModel();
bool findByOmName = all_dump_model.find(om_name_) != all_dump_model.end();
bool findByModelName = all_dump_model.find(name_) != all_dump_model.end();
bool dump_l1fusion_op = (all_dump_model.find(ge::DUMP_ALL_MODEL) != all_dump_model.end()) ||
findByOmName || findByModelName;
if (dump_l1fusion_op) {
// malloc 2M for dump l1fusion op
GE_CHK_RT_RET(rtMalloc(&l1_fusion_addr_, kDumpL1FusionOpMByteSize, RT_MEMORY_DDR));
// send l1fusion dump addr to rts
GE_CHK_RT_RET(rtDumpAddrSet(rt_model_handle_, l1_fusion_addr_, kDumpL1FusionOpMByteSize, kDumpFlagOfL1Fusion));
}
/// In zero copy model, if a aicpu operator is connected to the first or last layer, before model execution,
/// the aicpu opertor needs to destroy history record, and update operator memory address.
/// The model with specified aicpu operators is only marked here, and destruction is in ModelManager::ExecuteModel().
need_destroy_aicpu_kernel_ = IsAicpuKernelConnectSpecifiedLayer();
(void)ge::AttrUtils::GetListStr(ge_model_, ATTR_MODEL_OUT_NODES_NAME, out_node_name_);
string fp_ceiling_mode;
if (ge::AttrUtils::GetStr(ge_model_, ATTR_FP_CEILING_MODE, fp_ceiling_mode)) {
GELOGI("Get attr ATTR_FP_CEILING_MODE from model, value is %s.", fp_ceiling_mode.c_str());
// mode 0: Do not perform saturation processing. By default, IEEE754 is used.
GE_CHK_RT_RET(rtSetCtxINFMode((fp_ceiling_mode != "0")));
}
// collect profiling for ge
auto &profiling_manager = ProfilingManager::Instance();
if (profiling_manager.ProfilingModelLoadOn()) {
Status p_ret = ReportProfilingData(!profiling_manager.IsAclApiMode());
if (p_ret != SUCCESS) {
GELOGE(p_ret, "Report profiling data failed.");
return p_ret;
}
}
Shrink();
GELOGI("Davinci model init success.");
return ret;
}
Status DavinciModel::ReportProfilingData(bool check_device) {
std::vector<ComputeGraphDescInfo> compute_graph_desc_info;
Status ret = GetComputeGraphInfo(compute_graph_desc_info);
if (ret != SUCCESS) {
GELOGE(ret, "GetComputeGraphInfo failed.");
return ret;
}
ProfilingManager::Instance().ReportProfilingData(model_id_, GetTaskDescInfo(), compute_graph_desc_info, check_device);
GE_CHK_STATUS(SinkModelProfile(), "Sink model profiler failed.");
op_list_.clear();
return SUCCESS;
}
///
/// @ingroup ge
/// @brief Travel all nodes and determine if destruction is required.
/// @return bool
///
bool DavinciModel::IsAicpuKernelConnectSpecifiedLayer() {
Graph graph = ge_model_->GetGraph();
ComputeGraphPtr compute_graph = GraphUtils::GetComputeGraph(graph);
auto all_nodes = compute_graph->GetAllNodes();
for (auto &node : all_nodes) {
GE_IF_BOOL_EXEC(node == nullptr, continue);
OpDescPtr op_desc = node->GetOpDesc();
GE_IF_BOOL_EXEC(op_desc == nullptr, continue);
int64_t imply_type = -1;
(void)ge::AttrUtils::GetInt(op_desc, ATTR_NAME_IMPLY_TYPE, imply_type);
if (imply_type != static_cast<int64_t>(domi::ImplyType::AI_CPU)) {
continue;
}
GELOGD("Current operator imply type is %ld, name is %s.", imply_type, op_desc->GetName().c_str());
for (auto &in_data_anchor : node->GetAllInDataAnchors()) {
GE_IF_BOOL_EXEC(in_data_anchor == nullptr, continue);
auto peer_out_data_anchor = in_data_anchor->GetPeerOutAnchor();
GE_IF_BOOL_EXEC(peer_out_data_anchor == nullptr, continue);
auto peer_node = peer_out_data_anchor->GetOwnerNode();
GE_IF_BOOL_EXEC(peer_node == nullptr, continue);
auto peer_op_desc = peer_node->GetOpDesc();
GE_IF_BOOL_EXEC(peer_op_desc == nullptr, continue);
if (IsDataOp(peer_op_desc->GetType())) {
GELOGI("Mark specified aicpu operator connected to data.");
return true;
}
}
for (auto &out_data_anchor : node->GetAllOutDataAnchors()) {
GE_IF_BOOL_EXEC(out_data_anchor == nullptr, continue);
auto peer_in_data_anchors = out_data_anchor->GetPeerInDataAnchors();
for (auto &peer_in_data_anchor : peer_in_data_anchors) {
GE_IF_BOOL_EXEC(peer_in_data_anchor == nullptr, continue);
auto peer_node = peer_in_data_anchor->GetOwnerNode();
GE_IF_BOOL_EXEC(peer_node == nullptr, continue);
auto peer_op_desc = peer_node->GetOpDesc();
GE_IF_BOOL_EXEC(peer_op_desc == nullptr, continue);
if (peer_op_desc->GetType() == NETOUTPUT) {
GELOGI("Mark specified aicpu operator connected to netoutput.");
return true;
}
}
}
}
return false;
}
Status DavinciModel::UpdateSessionId(uint64_t session_id) {
GE_CHECK_NOTNULL(ge_model_);
if (!AttrUtils::SetInt(ge_model_, MODEL_ATTR_SESSION_ID, static_cast<int64_t>(session_id))) {
GELOGW("Set attr[%s] failed in updating session_id.", MODEL_ATTR_SESSION_ID.c_str());
}
GELOGD("Update session id: %lu.", session_id);
return SUCCESS;
}
///
/// @ingroup ge
/// @brief Travel all nodes and do some init.
/// @param [in] compute_graph: ComputeGraph to load.
/// @return Status
///
Status DavinciModel::InitNodes(const ComputeGraphPtr &compute_graph) {
uint32_t data_op_index = 0;
GE_TIMESTAMP_CALLNUM_START(LoadTBEKernelBinToOpDesc);
GE_TIMESTAMP_CALLNUM_START(InitTbeHandle);
typedef Status (DavinciModel::*OpDescCall)(const OpDescPtr &);
static std::map<std::string, OpDescCall> op_desc_handle = {
{VARIABLE, &DavinciModel::InitVariable},
{CONSTANTOP, &DavinciModel::InitConstant},
{STREAMACTIVE, &DavinciModel::InitStreamActive},
{STREAMSWITCH, &DavinciModel::InitStreamSwitch},
{STREAMSWITCHN, &DavinciModel::InitStreamSwitchN},
{LABELSET, &DavinciModel::InitLabelSet},
{CASE, &DavinciModel::InitCase},
};
GE_CHK_STATUS_RET(InitInputOutputForDynamic(compute_graph), "InitInputOutputForDynamic failed.");
map<uint32_t, OpDescPtr> data_by_index;
auto nodes = compute_graph->GetAllNodes();
const CustAICPUKernelStore &aicpu_kernel_store = ge_model_->GetCustAICPUKernelStore();
for (size_t i = 0; i < nodes.size(); i++) {
auto node = nodes.at(i);
auto op_desc = node->GetOpDesc();
if (op_desc == nullptr) {
GELOGE(PARAM_INVALID, "op_desc is null.");
return PARAM_INVALID;
}
op_list_[op_desc->GetId()] = op_desc;
GE_TIMESTAMP_RESTART(LoadTBEKernelBinToOpDesc);
aicpu_kernel_store.LoadCustAICPUKernelBinToOpDesc(op_desc);
GE_TIMESTAMP_ADD(LoadTBEKernelBinToOpDesc);
if (IsDataOp(op_desc->GetType())) {
if (InitDataOp(node, data_op_index, data_by_index) != SUCCESS) {
GELOGE(PARAM_INVALID, "Data init failed, Name: %s", op_desc->GetName().c_str());
return PARAM_INVALID;
}
data_dumper_.SaveDumpInput(node);
continue;
}
if (op_desc->GetType() == NETOUTPUT) {
if (InitNetOutput(node) != SUCCESS) {
GELOGE(PARAM_INVALID, "NetOutput init failed, Name: %s", op_desc->GetName().c_str());
return PARAM_INVALID;
}
continue;
}
auto it = op_desc_handle.find(op_desc->GetType());
if (it != op_desc_handle.end()) {
if ((this->*it->second)(op_desc) != SUCCESS) {
GELOGE(PARAM_INVALID, "NetOutput init failed, Name: %s", op_desc->GetName().c_str());
return PARAM_INVALID;
}
continue;
}
// for dynamic shape with control flow
SetLabelForDynamic(node);
if (IsNoTaskAndDumpNeeded(op_desc)) {
GELOGD("node[%s] without task, and save op_desc and addr for dump", op_desc->GetName().c_str());
const RuntimeParam &rts_param = GetRuntimeParam();
const vector<void *> input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc);
const vector<void *> output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc);
const vector<void *> workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc);
vector<void *> tensor_device_addrs;
tensor_device_addrs.insert(tensor_device_addrs.end(), input_data_addrs.begin(), input_data_addrs.end());
tensor_device_addrs.insert(tensor_device_addrs.end(), output_data_addrs.begin(), output_data_addrs.end());
tensor_device_addrs.insert(tensor_device_addrs.end(), workspace_data_addrs.begin(), workspace_data_addrs.end());
void *addr = nullptr;
auto size = kAddrLen * tensor_device_addrs.size();
GE_CHK_RT_RET(rtMalloc(&addr, size, RT_MEMORY_HBM));
rtError_t rt_ret = rtMemcpy(addr, size, tensor_device_addrs.data(), size, RT_MEMCPY_HOST_TO_DEVICE);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "rtMemcpy error, ret: 0x%X", rt_ret);
GE_CHK_RT(rtFree(addr));
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
saved_task_addrs_.emplace(op_desc, addr);
}
GE_TIMESTAMP_RESTART(InitTbeHandle);
uint32_t run_mode = static_cast<uint32_t>(domi::ImplyType::INVALID);
if (AttrUtils::GetInt(op_desc, ATTR_NAME_IMPLY_TYPE, run_mode) &&
run_mode == static_cast<uint32_t>(domi::ImplyType::TVM)) {
// Skip no_task operator, such as concat and split.
bool attr_notask = false;
bool get_attr_notask_flag = ge::AttrUtils::GetBool(op_desc, ATTR_NAME_NOTASK, attr_notask);
GE_IF_BOOL_EXEC(get_attr_notask_flag && attr_notask,
GELOGI("Node[name:%s, type:%s] does not generate task, skip initialization.",
op_desc->GetName().c_str(), op_desc->GetType().c_str());
continue;);
Status status = InitTbeHandle(op_desc);
if (status != SUCCESS) {
GELOGE(status, "TBE init failed. %s", op_desc->GetName().c_str());
return status;
}
}
GE_TIMESTAMP_ADD(InitTbeHandle);
}
AdjustDataOpList(data_by_index);
GE_TIMESTAMP_CALLNUM_END(LoadTBEKernelBinToOpDesc, "GraphLoader::LoadTBEKernelBinToOpDesc.");
GE_TIMESTAMP_CALLNUM_END(InitTbeHandle, "GraphLoader::InitTbeHandle.");
return SUCCESS;
}
Status DavinciModel::InitInputOutputForDynamic(const ComputeGraphPtr &compute_graph) {
if (!known_node_) return SUCCESS;
// for dynamic shape
auto direct_nodes = compute_graph->GetDirectNode();
for (size_t i = 0; i < direct_nodes.size(); i++) {
auto node = direct_nodes.at(i);
auto op_desc = node->GetOpDesc();
if (op_desc == nullptr) {
GELOGE(PARAM_INVALID, "op_desc is null.");
return PARAM_INVALID;
}
if (IsDataOp(op_desc->GetType())) {
GELOGD("init data op %s", op_desc->GetName().c_str());
data_op_list_.push_back(op_desc);
}
if (op_desc->GetType() == NETOUTPUT) {
GELOGD("init netouput op %s", op_desc->GetName().c_str());
output_op_list_.push_back(op_desc);
}
}
return SUCCESS;
}
void DavinciModel::SetLabelForDynamic(const NodePtr &node) {
if (known_node_ && node->GetOpDesc()->GetType() == LABELSWITCHBYINDEX) {
for (auto &in_data_anchor : node->GetAllInDataAnchors()) {
auto peer_out_data_anchor = in_data_anchor->GetPeerOutAnchor();
if (peer_out_data_anchor != nullptr) {
string tensor_name = node->GetName();
auto peer_node = peer_out_data_anchor->GetOwnerNode();
(void)AttrUtils::SetStr(peer_node->GetOpDesc(), ATTR_DYNAMIC_SHAPE_FIXED_ADDR, tensor_name);
(void)AttrUtils::SetInt(peer_node->GetOpDesc(), ATTR_DYNAMIC_SHAPE_FIXED_ADDR_INDEX, 0);
tensor_name_to_peer_output_index_[tensor_name] = 0;
}
}
}
}
/// @ingroup ge
/// @brief Data Op Initialize.
/// @param [in] NodePtr: Data Op.
/// @param [in/out] data_op_index: NetOutput addr size info.
/// @return Status
Status DavinciModel::InitDataOp(const NodePtr &node, uint32_t &data_op_index, map<uint32_t, OpDescPtr> &data_by_index) {
// op_desc Checked by Init: Data, valid.
auto op_desc = node->GetOpDesc();
if (known_node_) {
return SUCCESS;
}
uint32_t parent_index = 0; // Ignore subgraph Data Node.
if (AttrUtils::GetInt(op_desc, ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
GELOGI("Init zero copy by subgraph Data node: %s.", op_desc->GetName().c_str());
return InitInputBatchLabel(node);
}
data_op_list_.push_back(op_desc);
// Make information for copy input data.
const vector<int64_t> output_size_list = ModelUtils::GetOutputSize(op_desc);
const vector<void *> virtual_addr_list = ModelUtils::GetOutputDataAddrs(runtime_param_, op_desc);
const vector<int64_t> output_offset_list = op_desc->GetOutputOffset();
if (output_offset_list.size() != virtual_addr_list.size()) {
GELOGE(PARAM_INVALID, "virtual_addr size:%zu should be equal to offset size:%zu.", virtual_addr_list.size(),
output_offset_list.size());
return PARAM_INVALID;
}
auto data_index = data_op_index;
if (AttrUtils::GetInt(op_desc, ATTR_NAME_INDEX, data_index)) {
GELOGI("ge_train: get new index %u, old %u", data_index, data_op_index);
}
bool fusion_flag = false;
ZeroCopyOffset zero_copy_offset;
Status ret = zero_copy_offset.InitInputDataInfo(output_size_list, virtual_addr_list, op_desc, fusion_flag);
if (ret != SUCCESS) {
GELOGE(PARAM_INVALID, "InitDataInfo of input_info %s failed.", op_desc->GetName().c_str());
return PARAM_INVALID;
}
new_input_data_info_[data_index] = zero_copy_offset;
data_by_index[data_index] = op_desc;
for (size_t index = 0; index < virtual_addr_list.size(); ++index) {
void *addr = virtual_addr_list.at(index);
if (new_input_outside_addrs_.find(addr) != new_input_outside_addrs_.end()) {
continue;
}
zero_copy_offset.SetInputOutsideAddrs(output_offset_list, addr, index, fusion_flag, real_virtual_addrs_);
new_input_outside_addrs_[addr] = zero_copy_offset;
}
GELOGI("SetInputOutsideAddr success.");
data_op_index++;
if (InitInputZeroCopyNodes(node) != SUCCESS) {
GELOGE(PARAM_INVALID, "Input zero copy nodes init failed!");
return PARAM_INVALID;
}
return SUCCESS;
}
///
/// @ingroup ge
/// @brief Sort Data op list by index.
/// @param [in] data_by_index: map of Data Op.
/// @return
///
void DavinciModel::AdjustDataOpList(const map<uint32_t, OpDescPtr> &data_by_index) {
if (data_by_index.size() != data_op_list_.size()) {
GELOGW("Data map size: %zu, Data list size: %zu.", data_by_index.size(), data_op_list_.size());
return;
}
data_op_list_.clear();
for (auto &item : data_by_index) {
data_op_list_.emplace_back(item.second);
}
}
///
/// @ingroup ge
/// @brief input zero copy node Initialize.
/// @param [in] NodePtr: Data Op.
/// @return Status
///
Status DavinciModel::InitInputZeroCopyNodes(const NodePtr &node) {
auto out_data_anchor = node->GetOutDataAnchor(kDataIndex);
if (out_data_anchor == nullptr) {
GELOGE(FAILED, "Out data anchor is nullptr");
return FAILED;
}
for (auto &peer_in_data_anchor : out_data_anchor->GetPeerInDataAnchors()) {
auto node = peer_in_data_anchor->GetOwnerNode();
auto op_desc = node->GetOpDesc();
if (op_desc == nullptr) {
GELOGE(FAILED, "Op desc is nullptr");
return FAILED;
}
string batch_label;
(void)ge::AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
if (batch_label.empty()) {
batch_label = kDefaultBatchLable;
}
if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
zero_copy_op_id_batch_label_.emplace(pair<int64_t, string>(op_desc->GetId(), batch_label));
GELOGD("Init input zero copy nodes success, op name:%s, op id: %ld, batch label: %s.", op_desc->GetName().c_str(),
op_desc->GetId(), batch_label.c_str());
}
}
return SUCCESS;
}
bool DavinciModel::IsGetNextSinkDynamic(const OpDescPtr &op_desc) {
bool getnext_sink_dynamic = false;
if (ge::AttrUtils::GetBool(op_desc, ATTR_GETNEXT_SINK_DYNMAIC, getnext_sink_dynamic) && getnext_sink_dynamic) {
GELOGI("ATTR_GETNEXT_SINK_DYNMAIC has been set and is true.");
return true;
}
return false;
}
/// @ingroup ge
/// @brief NetOutput Op Initialize.
/// @param [in] NodePtr: NetOutput Op.
/// @return Status
Status DavinciModel::InitNetOutput(const NodePtr &node) {
// node->GetOpDesc Checked by Init: NetOutput, valid.
auto op_desc = node->GetOpDesc();
// excludes the function op sub graph, e.g. case,if
if (known_node_) {
return SUCCESS;
}
ComputeGraphPtr owner_graph = node->GetOwnerComputeGraph();
GE_CHECK_NOTNULL(owner_graph);
if (owner_graph->GetParentGraph() != nullptr) {
GELOGI("Init zero copy by subgraph NetOutput node: %s.", op_desc->GetName().c_str());
op_list_.erase(op_desc->GetId());
return InitOutputBatchLabel(node);
}
output_op_list_.push_back(op_desc);
// Make information for copy output data.
const vector<int64_t> input_size_list = ModelUtils::GetInputSize(op_desc);
const vector<void *> virtual_addr_list = ModelUtils::GetInputDataAddrs(runtime_param_, op_desc);
const vector<int64_t> input_offset_list = op_desc->GetInputOffset();
GE_IF_BOOL_EXEC(input_offset_list.size() != virtual_addr_list.size(),
GELOGE(PARAM_INVALID, "virtual_addr size should be equal to offset size."); return PARAM_INVALID;);
if (input_size_list.empty() && virtual_addr_list.empty()) {
GELOGI("NetOutput[%s] is empty.", op_desc->GetName().c_str());
return SUCCESS;
}
if (input_size_list.empty() || input_size_list.size() != virtual_addr_list.size()) {
GELOGE(PARAM_INVALID, "NetOutput[%s] init failed: Input size is %zu, Input addr is %zu", op_desc->GetName().c_str(),
input_size_list.size(), virtual_addr_list.size());
return PARAM_INVALID;
}
size_t num = new_output_data_info_.size();
bool fusion_flag = false;
size_t input_count = input_size_list.size();
is_getnext_sink_dynamic_ = false;
if (IsGetNextSinkDynamic(op_desc)) {
input_count = input_size_list.size() - kGetDynamicDimsCount;
is_getnext_sink_dynamic_ = true;
}
for (size_t idx = 0; idx < input_count; ++idx) {
ZeroCopyOffset zero_copy_offset;
Status ret = zero_copy_offset.InitOutputDataInfo(input_size_list, virtual_addr_list, op_desc, idx, fusion_flag);
GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(PARAM_INVALID, "InitDataInfo of input_info %s failed.",
op_desc->GetName().c_str()); return PARAM_INVALID;);
new_output_data_info_[num + idx] = zero_copy_offset;
void *addr = virtual_addr_list.at(idx);
int64_t input_offset = input_offset_list.at(idx);
vector<void *> tensor_addrs;
zero_copy_offset.SetOutputOutsideAddrs(input_offset, fusion_flag, addr, tensor_addrs);
auto rslt = new_output_outside_addrs_.insert(std::pair<void *, ZeroCopyOffset>(addr, zero_copy_offset));
if (!rslt.second) {
GELOGI("same output_tensor_addr %p to different input_tensor of %s", addr, op_desc->GetName().c_str());
DisableZeroCopy(addr);
}
for (size_t i = 0; i < tensor_addrs.size(); ++i) {
void *real_addr = tensor_addrs.at(i);
DisableZeroCopy(real_addr);
real_virtual_addrs_.insert(real_addr);
}
GELOGI("SetOutputOutsideAddr success.");
}
GE_IF_BOOL_EXEC(InitOutputZeroCopyNodes(node) != SUCCESS,
GELOGE(PARAM_INVALID, "Output zero copy nodes init failed!"); return PARAM_INVALID;);
GetAllGearsInfo(node);
if (is_getnext_sink_dynamic_) {
GE_IF_BOOL_EXEC(GetGetDynamicDimsNodeInfo(node) != SUCCESS,
GELOGE(PARAM_INVALID, "Failed to get info of getdynamicdims node."); return PARAM_INVALID;);
}
if (is_online_infer_dynamic_) {
GE_IF_BOOL_EXEC(GetGearAndRealOutSizeInfo(input_count, node) != SUCCESS,
GELOGE(PARAM_INVALID, "Failed to get gear and real out size info."); return PARAM_INVALID;);
GE_IF_BOOL_EXEC(GetGearAndRealOutShapeInfo(input_count, op_desc) != SUCCESS,
GELOGE(PARAM_INVALID, "Failed to get gear and real out shape info."); return PARAM_INVALID;);
}
GELOGI("DavinciModel::InitNetoutput success.");
return SUCCESS;
}
void DavinciModel::GetAllGearsInfo(const NodePtr &node) {
is_online_infer_dynamic_ = false;
all_gears_info_.clear();
std::string shapes;
(void) AttrUtils::GetStr(node->GetOpDesc(), ATTR_ALL_GEARS_INFO, shapes);
if (!shapes.empty()) {
is_online_infer_dynamic_ = true;
std::vector<std::string> shape_strs = ge::StringUtils::Split(shapes, ';');
for (const auto &shape_str : shape_strs) {
if (shape_str.empty()) {
continue;
}
std::vector<int64_t> gear_info;
std::vector<std::string> dims = ge::StringUtils::Split(shape_str, ',');
for (const auto &dim : dims) {
if (dim.empty()) {
continue;
}
gear_info.emplace_back(std::strtol(dim.c_str(), nullptr, kDecimal));
}
if (!gear_info.empty()) {
all_gears_info_.emplace_back(gear_info);
GELOGD("Init all gears info from %s, gaer info is %s.", node->GetName().c_str(),
formats::JoinToString(gear_info).c_str());
}
}
}
}
Status DavinciModel::GetGetDynamicDimsNodeInfo(const NodePtr &node) {
GE_CHECK_NOTNULL(node->GetOpDesc());
size_t input_count = node->GetAllInDataAnchors().size();
GELOGI("input_anchor count of %s is %zu.", node->GetName().c_str(), input_count);
size_t get_dynamic_dims_index = input_count - kGetDynamicDimsCount;
auto in_anchor = node->GetAllInDataAnchors().at(get_dynamic_dims_index);
auto peer_out_anchor = in_anchor->GetPeerOutAnchor();
if (peer_out_anchor == nullptr) {
GELOGE(PARAM_INVALID, "Out anchor of getdynmaicdims node should not be nullptr.");
return PARAM_INVALID;
}
auto peer_node = peer_out_anchor->GetOwnerNode();
auto op_desc = peer_node->GetOpDesc();
GE_CHECK_NOTNULL(op_desc);
if (op_desc->GetName() == kGetDynamicDimsName && op_desc->GetType() == GETDYNAMICDIMS) {
GELOGD("Start get info of %s.", op_desc->GetName().c_str());
auto input_addr = ModelUtils::GetInputDataAddrs(runtime_param_, node->GetOpDesc());
auto input_size = ModelUtils::GetInputSize(node->GetOpDesc());
if (input_addr.empty() || input_size.empty()) {
GELOGE(PARAM_INVALID, "Not set output of %s", op_desc->GetName().c_str());
return PARAM_INVALID;
}
auto input_desc = node->GetOpDesc()->GetInputDescPtr(get_dynamic_dims_index);
GE_CHECK_NOTNULL(input_desc);
if (input_desc->GetShape().GetDims().empty()) {
GELOGE(PARAM_INVALID, "Not set output desc shape of %s.", op_desc->GetName().c_str());
return PARAM_INVALID;
}
netoutput_last_input_addr_ = input_addr[get_dynamic_dims_index];
netoutput_last_input_size_ = input_size[get_dynamic_dims_index];
shape_of_cur_dynamic_dims_ = input_desc->GetShape().GetDims().at(0);
GELOGD("Shape of cur dynamic dims is %zu, size is %ld, addr is %p.", shape_of_cur_dynamic_dims_,
netoutput_last_input_size_, netoutput_last_input_addr_);
}
return SUCCESS;
}
Status DavinciModel::GetGearAndRealOutSizeInfo(size_t input_count, const NodePtr &node) {
GELOGD("Start get gear and real output size info of %s, input count is %zu.", node->GetName().c_str(), input_count);
merge_nodes_gear_and_real_out_size_info_.clear();
for (size_t idx = 0; idx < input_count; ++idx) {
auto in_anchor = node->GetAllInDataAnchors().at(idx);
auto peer_out_anchor = in_anchor->GetPeerOutAnchor();
if (peer_out_anchor == nullptr) {
continue;
}
auto peer_node = peer_out_anchor->GetOwnerNode();
auto op_desc = peer_node->GetOpDesc();
GE_CHECK_NOTNULL(op_desc);
if ((peer_node->GetType() == MERGE) && (op_desc->HasAttr(ATTR_INSERT_BY_MBATCH))) {
if (GetRealOutputSizeOfMerge(idx, peer_node) != SUCCESS) {
GELOGE(PARAM_INVALID, "Get real output size of %s failed.", peer_node->GetName().c_str());
return PARAM_INVALID;
}
}
}
return SUCCESS;
}
Status DavinciModel::GetRealOutputSizeOfMerge(size_t input_index, const NodePtr &merge_node) {
GELOGD("Start get output size of %s, which is %zu input to netoutput.", merge_node->GetName().c_str(), input_index);
std::map<vector<int64_t>, int64_t> gear_and_real_out_size_info;
for (auto &in_anchor : merge_node->GetAllInDataAnchors()) {
auto peer_out_anchor = in_anchor->GetPeerOutAnchor();
if (peer_out_anchor == nullptr) {
continue;
}
auto in_node = peer_out_anchor->GetOwnerNode();
GELOGD("Input node of merge is %s.", in_node->GetName().c_str());
auto op_desc = in_node->GetOpDesc();
GE_CHECK_NOTNULL(op_desc);
string batch_label;
if (AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label)) {
size_t batch_index = static_cast<size_t>(stoi(batch_label.substr(batch_label.rfind('_') + 1)));
GELOGD("Batch index of %s is %zu.", op_desc->GetName().c_str(), batch_index);
if (batch_index > all_gears_info_.size()) {
GELOGE(PARAM_INVALID, "The value of ATTR_NAME_BATCH_LABEL is invalid.");
return PARAM_INVALID;
}
const vector<int64_t> output_size_list = ModelUtils::GetOutputSize(op_desc);
int output_index = ge::AnchorUtils::GetIdx(peer_out_anchor);
auto tensor_desc = op_desc->GetOutputDescPtr(output_index);
GE_CHECK_NOTNULL(tensor_desc);
int64_t data_size = 0;
if (TensorUtils::GetTensorSizeInBytes(*tensor_desc, data_size) != GRAPH_SUCCESS) {
GELOGE(FAILED, "Get tensor size in bytes failed.");
return FAILED;
}
gear_and_real_out_size_info[all_gears_info_[batch_index]] = data_size;
GELOGD("Get real gear index is: %zu, gear info is %s, size is %ld, tensor size is %ld",
batch_index, formats::JoinToString(all_gears_info_[batch_index]).c_str(),
output_size_list[output_index], data_size);
}
}
merge_nodes_gear_and_real_out_size_info_[input_index] = gear_and_real_out_size_info;
return SUCCESS;
}
Status DavinciModel::GetGearAndRealOutShapeInfo(size_t input_count, const OpDescPtr &op_desc) {
GELOGD("Start to get dynamic output dims of %s.", op_desc->GetName().c_str());
merge_nodes_gear_and_real_out_shape_info_.clear();
std::vector<std::string> dynamic_output_shape_info;
if (!AttrUtils::GetListStr(op_desc, ATTR_NAME_DYNAMIC_OUTPUT_DIMS, dynamic_output_shape_info)) {
GELOGD("Can not get dynamic output dims attr");
return SUCCESS;
}
GELOGI("Dynamic output shape info is %s", formats::JoinToString(dynamic_output_shape_info).c_str());
std::vector<vector<int64_t>> dynamic_output_shape;
ParseDynamicOutShape(dynamic_output_shape_info, dynamic_output_shape);
// idx: input_index to netoutput
for (size_t idx = 0; idx < input_count; ++idx) {
std::map<vector<int64_t>, vector<int64_t>> gear_and_real_out_shape_info;
for (auto &it : dynamic_output_shape) {
auto gear_index = static_cast<size_t>(it[0]);
if (gear_index > all_gears_info_.size()) {
GELOGE(PARAM_INVALID, "The value of cur index: %zu is invalid.", static_cast<size_t>(it[0]));
return PARAM_INVALID;
}
if (static_cast<size_t>(it[1]) == idx) {
vector<int64_t> output_shape;
for (size_t i = 2; i < it.size(); ++i) {
output_shape.emplace_back(it[i]);
}
gear_and_real_out_shape_info[all_gears_info_[gear_index]] = output_shape;
GELOGD("Get real gear index is: %zu, gear info is %s, output shape is %s.",
gear_index, formats::JoinToString(all_gears_info_[gear_index]).c_str(),
formats::JoinToString(output_shape).c_str());
}
}
merge_nodes_gear_and_real_out_shape_info_[idx] = gear_and_real_out_shape_info;
}
return SUCCESS;
}
void DavinciModel::ParseDynamicOutShape(const std::vector<std::string> &str_info,
std::vector<vector<int64_t>> &vec_info) {
for (size_t i = 0; i < str_info.size(); ++i) {
std::vector<int64_t> shape;
std::vector<std::string> dims = ge::StringUtils::Split(str_info[i], ',');
for (const auto &dim : dims) {
if (dim.empty()) {
continue;
}
shape.emplace_back(std::strtol(dim.c_str(), nullptr, kDecimal));
}
GELOGI("Shape from attr is %s.", formats::JoinToString(shape).c_str());
vec_info.emplace_back(shape);
}
}
///
/// @ingroup ge
/// @brief output zero copy node Initialize.
/// @param [in] NodePtr: netoutput Op.
/// @return Status
///
Status DavinciModel::InitOutputZeroCopyNodes(const NodePtr &node) {
set<NodePtr> nodes_need_record;
for (auto &in_data_anchor : node->GetAllInDataAnchors()) {
auto peer_out_data_anchor = in_data_anchor->GetPeerOutAnchor();
if (peer_out_data_anchor == nullptr) {
continue;
}
auto peer_node = peer_out_data_anchor->GetOwnerNode();
nodes_need_record.emplace(peer_node);
// Merge node output multiplexed input, upstream nodes need to be considered in multiple batch scenarios
if (peer_node->GetType() == MERGE) {
for (const auto &merge_peer_in_data_anchor : peer_node->GetAllInDataAnchors()) {
auto merge_peer_out_data_anchor = merge_peer_in_data_anchor->GetPeerOutAnchor();
if (merge_peer_out_data_anchor == nullptr) {
continue;
}
auto merge_peer_node = merge_peer_out_data_anchor->GetOwnerNode();
nodes_need_record.emplace(merge_peer_node);
}
} else {
for (const auto &other_in_data_anchor : peer_out_data_anchor->GetPeerInDataAnchors()) {
auto other_in_node = other_in_data_anchor->GetOwnerNode();
if (other_in_node->GetType() != NETOUTPUT) {
nodes_need_record.emplace(other_in_node);
}
}
}
}
for (const auto &node_need_record : nodes_need_record) {
auto op_desc = node_need_record->GetOpDesc();
GE_CHECK_NOTNULL(op_desc);
string batch_label;
(void)ge::AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
if (batch_label.empty()) {
batch_label = kDefaultBatchLable;
}
if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
zero_copy_op_id_batch_label_.emplace(pair<int64_t, string>(op_desc->GetId(), batch_label));
GELOGD("Init Output zero copy nodes success, op name:%s, op id: %ld, batch label: %s.",
op_desc->GetName().c_str(), op_desc->GetId(), batch_label.c_str());
}
}
return SUCCESS;
}
///
/// @ingroup ge
/// @brief input zero copy node Initialize.
/// @param [in] NodePtr: Data Op.
/// @return Status
///
Status DavinciModel::InitInputBatchLabel(const NodePtr &node) {
string batch_label;
if (!AttrUtils::GetStr(node->GetOpDesc(), ATTR_NAME_BATCH_LABEL, batch_label)) {
return SUCCESS; // Not Multi-batch.
}
const auto &out_data_anchor = node->GetOutDataAnchor(kDataIndex);
GE_CHECK_NOTNULL(out_data_anchor);
for (const auto &peer_in_data_anchor : out_data_anchor->GetPeerInDataAnchors()) {
const auto &node = peer_in_data_anchor->GetOwnerNode();
const auto &op_desc = node->GetOpDesc();
GE_CHECK_NOTNULL(op_desc);
if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
zero_copy_op_id_batch_label_[op_desc->GetId()] = batch_label;
GELOGD("Init input zero copy nodes success, op name: %s, op id: %ld, batch label: %s", op_desc->GetName().c_str(),
op_desc->GetId(), batch_label.c_str());
}
}
return SUCCESS;
}
///
/// @ingroup ge
/// @brief output zero copy node Initialize for Case.
/// @param [in] NodePtr: netoutput Op.
/// @return Status
///
Status DavinciModel::InitOutputBatchLabel(const NodePtr &node) {
string batch_label;
if (!AttrUtils::GetStr(node->GetOpDesc(), ATTR_NAME_BATCH_LABEL, batch_label)) {
return SUCCESS; // Not Multi-batch.
}
for (const auto &in_data_anchor : node->GetAllInDataAnchors()) {
const auto &peer_out_data_anchor = in_data_anchor->GetPeerOutAnchor();
if (peer_out_data_anchor == nullptr) {
continue;
}
const auto &peer_node = peer_out_data_anchor->GetOwnerNode();
const auto &op_desc = peer_node->GetOpDesc();
GE_CHECK_NOTNULL(op_desc);
if (zero_copy_op_id_batch_label_.find(op_desc->GetId()) == zero_copy_op_id_batch_label_.end()) {
zero_copy_op_id_batch_label_[op_desc->GetId()] = batch_label;
GELOGD("Init Output zero copy nodes success, op name: %s, op id: %ld, batch label: %s",
op_desc->GetName().c_str(), op_desc->GetId(), batch_label.c_str());
}
}
return SUCCESS;
}
/// @ingroup ge
/// @brief LabelSet Op Initialize.
/// @param [in] op_desc: LabelSet Op descriptor.
/// @return Status
Status DavinciModel::InitLabelSet(const OpDescPtr &op_desc) {
uint32_t label_index = 0;
if (!AttrUtils::GetInt(op_desc, ATTR_NAME_LABEL_SWITCH_INDEX, label_index)) {
GELOGE(INTERNAL_ERROR, "InitLabelSet: %s attr [%s] not exist.", op_desc->GetName().c_str(),
ATTR_NAME_LABEL_SWITCH_INDEX.c_str());
return INTERNAL_ERROR;
}
if (label_index >= LabelNum()) {
GELOGE(INTERNAL_ERROR, "InitLabelSet: label index: %u >= label size: %zu.", label_index, LabelNum());
return INTERNAL_ERROR;
}
if (label_id_indication_.count(label_index) > 0) {
GELOGE(INTERNAL_ERROR, "InitLabelSet: %s label index: %u already used.", op_desc->GetName().c_str(), label_index);
return INTERNAL_ERROR;
}
rtStream_t stream = nullptr;
uint32_t stream_id = static_cast<uint32_t>(op_desc->GetStreamId());
if (stream_list_.size() == 1) {
stream = stream_list_[0];
} else if (stream_list_.size() > stream_id) {
stream = stream_list_[stream_id];
} else {
GELOGE(INTERNAL_ERROR, "InitLabelSet: stream index: %u >= stream size: %zu.", stream_id, stream_list_.size());
return INTERNAL_ERROR;
}
rtLabel_t rt_label = nullptr;
rtError_t rt_error = rtLabelCreateEx(&rt_label, stream);
if (rt_error != RT_ERROR_NONE || rt_label == nullptr) {
GELOGE(INTERNAL_ERROR, "InitLabelSet: %s create label failed, error=0x%x.", op_desc->GetName().c_str(), rt_error);
return INTERNAL_ERROR;
}
GELOGI("InitLabelSet: label[%u]=%p stream[%u]=%p.", label_index, rt_label, stream_id, stream);
label_id_indication_.insert(label_index);
label_list_[label_index] = rt_label;
return SUCCESS;
}
Status DavinciModel::InitVariable(const OpDescPtr &op_desc) {
variable_op_list_.push_back(op_desc);
return SUCCESS;
}
/// @ingroup ge
/// @brief ACL case, Load task list with queue.
/// @param [in] input_queue_ids: input queue ids from user, nums equal Data Op.
/// @param [in] output_queue_ids: input queue ids from user, nums equal NetOutput Op.
/// @return: 0 for success / others for failed
Status DavinciModel::SetQueIds(const std::vector<uint32_t> &input_queue_ids,
const std::vector<uint32_t> &output_queue_ids) {
if (input_queue_ids.empty() && output_queue_ids.empty()) {
GELOGE(GE_EXEC_MODEL_QUEUE_ID_INVALID, "Param is empty");
return GE_EXEC_MODEL_QUEUE_ID_INVALID;
}
input_queue_ids_ = input_queue_ids;
output_queue_ids_ = output_queue_ids;
return SUCCESS;
}
///
/// @ingroup ge
/// @brief ACL case, Load task list with queue.
/// @param [in] input_que_ids: input queue ids from user, nums equal Data Op.
/// @param [in] output_que_ids: input queue ids from user, nums equal NetOutput Op.
/// @return: 0 for success / others for failed
///
Status DavinciModel::LoadWithQueue() {
if (input_queue_ids_.empty() && output_queue_ids_.empty()) {
return SUCCESS;
}
if (input_queue_ids_.size() != new_input_data_info_.size()) {
GELOGE(GE_EXEC_MODEL_QUEUE_ID_INVALID, "Input queue ids not match model: input_queue=%zu input_data=%zu",
input_queue_ids_.size(), new_input_data_info_.size());
return GE_EXEC_MODEL_QUEUE_ID_INVALID;
}
if (output_queue_ids_.size() != new_output_data_info_.size()) {
GELOGE(GE_EXEC_MODEL_QUEUE_ID_INVALID, "Output queue ids not match model: output_queue=%zu output_data=%zu",
output_queue_ids_.size(), new_output_data_info_.size());
return GE_EXEC_MODEL_QUEUE_ID_INVALID;
}
GE_CHK_STATUS_RET(AddHeadStream(), "Add head stream failed.");
// Binding input_queue and Data Op.
GE_CHK_STATUS_RET(BindInputQueue(), "Launch bind input queue failed.");
GE_CHK_STATUS_RET(CpuTaskModelZeroCopy(input_mbuf_list_, new_input_outside_addrs_), "Launch zero copy failed.");
// Binding output_queue and NetOutput Op.
GE_CHK_STATUS_RET(BindOutputQueue(), "Launch bind output queue failed.");
GE_CHK_STATUS_RET(CpuTaskModelZeroCopy(output_mbuf_list_, new_output_outside_addrs_), "Launch zero copy failed.");
GE_CHK_STATUS_RET(CpuActiveStream(), "Launch active entry stream failed.");
GE_CHK_STATUS_RET(CpuWaitEndGraph(), "Launch wait end graph failed.");
GE_CHK_STATUS_RET(BindEnqueue(), "Launch enqueue failed.");
GE_CHK_STATUS_RET(CpuModelRepeat(), "Launch model repeat failed.");
return SUCCESS;
}
/// @ingroup ge
/// @brief queue schedule, Bind input queue to Data output address.
/// @return: 0 for success / others for failed
Status DavinciModel::BindInputQueue() {
// Caller checked: input_queue_ids_.size() == input_size_list_.size() != input_addr_list_.size()
for (size_t i = 0; i < input_queue_ids_.size(); ++i) {
auto it = new_input_data_info_.find(i);
if (it == new_input_data_info_.end()) {
GELOGE(FAILED, "Input not match: tensor num=%zu, Queue id index=%zu", new_input_data_info_.size(), i);
return FAILED;
}
uint32_t queue_id = input_queue_ids_[i];
if (it->second.GetDataInfo().empty()) {
GELOGE(INTERNAL_ERROR, "the %zu input_queue not set data_info.", i);
return INTERNAL_ERROR;
}
uint32_t data_size = static_cast<uint32_t>(it->second.GetDataInfo().at(0).first);
uintptr_t data_addr = reinterpret_cast<uintptr_t>(it->second.GetDataInfo().at(0).second);
GELOGI("BindInputToQueue: graph_%u index[%zu] queue id[%u] output addr[0x%lx] output size[%u]",
runtime_param_.graph_id, i, queue_id, data_addr, data_size);
rtError_t rt_ret = rtModelBindQueue(rt_model_handle_, queue_id, RT_MODEL_INPUT_QUEUE);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rtModelBindQueue failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
if (CpuModelDequeue(queue_id) != SUCCESS) {
return INTERNAL_ERROR;
}
}
return SUCCESS;
}
/// @ingroup ge
/// @brief definiteness queue schedule, bind input queue to task.
/// @param [in] queue_id: input queue id from user.
/// @return: 0 for success / others for failed
Status DavinciModel::CpuModelDequeue(uint32_t queue_id) {
GELOGI("Set CpuKernel model dequeue task enter.");
std::shared_ptr<CpuTaskModelDequeue> dequeue_task = MakeShared<CpuTaskModelDequeue>(rt_entry_stream_);
if (dequeue_task == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskModelDequeue task failed.");
return MEMALLOC_FAILED;
}
// Get DataOp Output address and bind to queue.
uintptr_t in_mbuf = 0;
Status status = dequeue_task->Init(queue_id, in_mbuf);
if (status != SUCCESS) {
return status;
}
cpu_task_list_.push_back(dequeue_task);
input_mbuf_list_.push_back(in_mbuf);
GELOGI("Set CpuKernel model dequeue task success.");
return SUCCESS;
}
Status DavinciModel::CpuTaskModelZeroCopy(std::vector<uintptr_t> &mbuf_list,
std::map<const void *, ZeroCopyOffset> &outside_addrs) {
GELOGI("Set CpuKernel model zero_copy task enter.");
std::shared_ptr<CpuTaskZeroCopy> zero_copy = MakeShared<CpuTaskZeroCopy>(rt_entry_stream_);
if (zero_copy == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskZeroCopy task failed.");
return MEMALLOC_FAILED;
}
// mdc zero_copy not support l2 fusion
Status status = zero_copy->Init(mbuf_list, outside_addrs);
if (status != SUCCESS) {
return status;
}
cpu_task_list_.push_back(zero_copy);
GELOGI("Set CpuKernel model zero_copy task success.");
return SUCCESS;
}
/// @ingroup ge
/// @brief queue schedule, bind output queue to NetOutput input address.
/// @return: 0 for success / others for failed
Status DavinciModel::BindOutputQueue() {
// Caller checked: input_queue_ids_.size() == input_size_list_.size() != input_addr_list_.size()
for (size_t i = 0; i < output_queue_ids_.size(); ++i) {
auto it = new_output_data_info_.find(i);
if (it == new_output_data_info_.end()) {
GELOGE(FAILED, "Output not match: tensor num=%zu, Queue id index=%zu", new_output_data_info_.size(), i);
return FAILED;
}
uint32_t queue_id = output_queue_ids_[i];
if (it->second.GetDataInfo().empty()) {
GELOGE(INTERNAL_ERROR, "the %zu output_queue not set data_info.", i);
return INTERNAL_ERROR;
}
uint32_t data_size = static_cast<uint32_t>(it->second.GetDataInfo().at(0).first);
uintptr_t data_addr = reinterpret_cast<uintptr_t>(it->second.GetDataInfo().at(0).second);
GELOGI("BindOutputToQueue: graph_%u index[%zu] queue id[%u] input addr[0x%lx] input size[%u]",
runtime_param_.graph_id, i, queue_id, data_addr, data_size);
rtError_t rt_ret = rtModelBindQueue(rt_model_handle_, queue_id, RT_MODEL_OUTPUT_QUEUE);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rtModelBindQueue failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
Status status = CpuModelPrepareOutput(data_addr, data_size);
if (status != SUCCESS) {
return status;
}
}
return SUCCESS;
}
/// @ingroup ge
/// @brief definiteness queue schedule, bind output queue to task.
/// @param [in] addr: NetOutput Op input tensor address.
/// @param [in] size: NetOutput Op input tensor size.
/// @return: 0 for success / others for failed
Status DavinciModel::CpuModelPrepareOutput(uintptr_t addr, uint32_t size) {
GELOGI("Set CpuKernel model enqueue task enter.");
if (input_mbuf_list_.empty()) {
GELOGE(FAILED, "Need input mbuf for fill output mbuf head info.");
return FAILED;
}
std::shared_ptr<CpuTaskPrepareOutput> prepare_output = MakeShared<CpuTaskPrepareOutput>(rt_entry_stream_);
if (prepare_output == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskPrepareOutput task failed.");
return MEMALLOC_FAILED;
}
uintptr_t out_mbuf = 0;
if (prepare_output->Init(addr, size, input_mbuf_list_.back(), out_mbuf) != SUCCESS) {
return FAILED;
}
cpu_task_list_.push_back(prepare_output);
output_mbuf_list_.push_back(out_mbuf);
GELOGI("Set CpuKernel model enqueue task success.");
return SUCCESS;
}
///
/// @ingroup ge
/// @brief definiteness queue schedule, active original model stream.
/// @return: 0 for success / others for failed
///
Status DavinciModel::CpuActiveStream() {
GELOGI("Set CpuKernel active stream task enter.");
std::shared_ptr<CpuTaskActiveEntry> active_entry = MakeShared<CpuTaskActiveEntry>(rt_entry_stream_);
if (active_entry == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskActiveEntry task failed.");
return MEMALLOC_FAILED;
}
Status status = active_entry->Init(rt_head_stream_);
if (status != SUCCESS) {
return status;
}
cpu_task_list_.push_back(active_entry);
GELOGI("Set CpuKernel active stream task success.");
return SUCCESS;
}
/// @ingroup ge
/// @brief definiteness queue schedule, wait for end graph.
/// @return: 0 for success / others for failed
Status DavinciModel::CpuWaitEndGraph() {
GELOGI("Set CpuKernel wait end graph task enter.");
std::shared_ptr<CpuTaskWaitEndGraph> wait_endgraph = MakeShared<CpuTaskWaitEndGraph>(rt_entry_stream_);
if (wait_endgraph == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskWaitEndGraph task failed.");
return MEMALLOC_FAILED;
}
Status status = wait_endgraph->Init(runtime_model_id_);
if (status != SUCCESS) {
return status;
}
cpu_task_list_.push_back(wait_endgraph);
GELOGI("Set CpuKernel wait end graph task success.");
return SUCCESS;
}
Status DavinciModel::BindEnqueue() {
for (size_t i = 0; i < output_queue_ids_.size(); ++i) {
auto it = new_output_data_info_.find(i);
if (it == new_output_data_info_.end()) {
GELOGE(FAILED, "Output not match: tensor num=%zu, Queue id index=%zu", new_output_data_info_.size(), i);
return FAILED;
}
uint32_t queue_id = output_queue_ids_[i];
if (CpuModelEnqueue(queue_id, output_mbuf_list_[i]) != SUCCESS) {
return INTERNAL_ERROR;
}
}
return SUCCESS;
}
Status DavinciModel::CpuModelEnqueue(uint32_t queue_id, uintptr_t out_mbuf) {
GELOGI("Set CpuKernel model enqueue task enter.");
std::shared_ptr<CpuTaskModelEnqueue> model_enqueue = MakeShared<CpuTaskModelEnqueue>(rt_entry_stream_);
if (model_enqueue == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskModelEnqueue task failed.");
return MEMALLOC_FAILED;
}
Status status = model_enqueue->Init(queue_id, out_mbuf);
if (status != SUCCESS) {
return status;
}
cpu_task_list_.push_back(model_enqueue);
GELOGI("Set CpuKernel model enqueue task enter.");
return SUCCESS;
}
/// @ingroup ge
/// @brief definiteness queue schedule, repeat run model.
/// @return: 0 for success / others for failed
Status DavinciModel::CpuModelRepeat() {
GELOGI("Set CpuKernel repeat task enter.");
std::shared_ptr<CpuTaskModelRepeat> model_repeat = MakeShared<CpuTaskModelRepeat>(rt_entry_stream_);
if (model_repeat == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskModelRepeat task failed.");
return MEMALLOC_FAILED;
}
Status status = model_repeat->Init(runtime_model_id_);
if (status != SUCCESS) {
return status;
}
cpu_task_list_.push_back(model_repeat);
GELOGI("Set CpuKernel repeat task success.");
return SUCCESS;
}
Status DavinciModel::GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc,
vector<InputOutputDescInfo> &output_desc) {
if ((data_op_list_.empty()) || (data_op_list_[0]->GetInputsSize()) != 1) {
GELOGI("data_op_list_ is empty or input_desc size is not 1.");
} else {
std::vector<uint32_t> input_formats;
GE_CHK_STATUS_RET(GetInputDescInfo(input_desc, input_formats), "get input desc info failed.");
}
std::vector<uint32_t> outputFormats;
GE_CHK_STATUS_RET(GetOutputDescInfo(output_desc, outputFormats), "get output desc info failed.");
return SUCCESS;
}
Status DavinciModel::GetInputOutputDescInfo(vector<InputOutputDescInfo> &input_desc,
vector<InputOutputDescInfo> &output_desc,
std::vector<uint32_t> &input_formats,
std::vector<uint32_t> &outputFormats) {
if ((data_op_list_.empty()) || (data_op_list_[0]->GetInputsSize()) != 1) {
GELOGE(FAILED, "OP List Pointer is null or input_desc size is not 1!");
return FAILED;
}
GE_CHK_STATUS_RET(GetInputDescInfo(input_desc, input_formats), "get input desc info failed");
GE_CHK_STATUS_RET(GetOutputDescInfo(output_desc, outputFormats), "get ouput desc info failed");
return SUCCESS;
}
///
/// @ingroup ge
/// @brief Get dynamic batch_info
/// @param [out] batch_info
/// @param [out] dynamic_type
/// @return execute result
///
Status DavinciModel::GetDynamicBatchInfo(std::vector<std::vector<int64_t>> &batch_info, int32_t &dynamic_type) const {
dynamic_type = dynamic_type_;
batch_info = batch_info_;
return SUCCESS;
}
///
/// @ingroup ge
/// @brief Get combined dynamic dims info
/// @param [out] batch_info
/// @return None
///
void DavinciModel::GetCombinedDynamicDims(std::vector<std::vector<int64_t>> &batch_info) const {
batch_info.clear();
batch_info = combined_batch_info_;
}
///
/// @ingroup ge
/// @brief Get user designate shape order
/// @param [out] user_input_shape_order
/// @return None
///
void DavinciModel::GetUserDesignateShapeOrder(std::vector<std::string> &user_input_shape_order) const {
user_input_shape_order.clear();
user_input_shape_order = user_designate_shape_order_;
}
///
/// @ingroup ge
/// @brief Get AIPP input info
/// @param [in] index
/// @param [out] aipp_info
/// @return execute result
///
Status DavinciModel::GetAIPPInfo(uint32_t index, AippConfigInfo &aipp_info) {
GE_CHK_BOOL_RET_STATUS(index < data_op_list_.size(), PARAM_INVALID, "Index %u is invalid.", index);
OpDescPtr data_op = data_op_list_[index];
if (!data_op->HasAttr(ATTR_NAME_AIPP)) {
GELOGW("GetAIPPInfo: there is not AIPP related with index %u.", index);
return GE_AIPP_NOT_EXIST;
}
std::unique_ptr<domi::AippOpParams> aipp_params(new (std::nothrow) domi::AippOpParams());
GE_CHECK_NOTNULL(aipp_params);
ge::GeAttrValue::NAMED_ATTRS aipp_attr;
GE_CHK_BOOL_RET_STATUS(AttrUtils::GetNamedAttrs(data_op, ATTR_NAME_AIPP, aipp_attr), GE_AIPP_NOT_EXIST,
"Data node do not contain param aipp!");
GE_CHK_STATUS_RET(OpUtils::ConvertAippParams(aipp_attr, aipp_params.get()), "get aipp params failed");
GELOGI("GetAIPPInfo: node data: %s, type: %s, current index: %u, current node related input rank: %u",
data_op->GetName().c_str(), data_op->GetType().c_str(), index, aipp_params->related_input_rank());
GE_CHK_STATUS_RET(AippUtils::ConvertAippParams2AippInfo(aipp_params.get(), aipp_info),
"convert aipp params to aipp config info failed");
return SUCCESS;
}
Status DavinciModel::GetAippType(uint32_t index, InputAippType &type, size_t &aipp_index) {
GE_CHK_BOOL_RET_STATUS(index < data_op_list_.size(), PARAM_INVALID, "Index %u is invalid.", index);
// Set default value
type = DATA_WITHOUT_AIPP;
aipp_index = 0xFFFFFFFF; // default invalid value
OpDescPtr data_op = data_op_list_[index];
GE_CHECK_NOTNULL(data_op);
if (!data_op->HasAttr(ATTR_DATA_RELATED_AIPP_MODE)) {
GELOGW("There is no aipp releated info with index %u.", index);
return SUCCESS;
}
std::string data_mode;
(void)AttrUtils::GetStr(data_op, ATTR_DATA_RELATED_AIPP_MODE, data_mode);
if (data_mode == "static_aipp") {
type = DATA_WITH_STATIC_AIPP;
} else if (data_mode == "dynamic_aipp") {
type = DATA_WITH_DYNAMIC_AIPP;
} else if (data_mode == "dynamic_aipp_conf") {
type = DYNAMIC_AIPP_NODE;
} else {
GELOGE(INTERNAL_ERROR, "The info of aipp releated info %s is invalid with index %u.", data_mode.c_str(), index);
return INTERNAL_ERROR;
}
if (type == DATA_WITH_DYNAMIC_AIPP) {
string releated_name;
(void)AttrUtils::GetStr(data_op, ATTR_DATA_AIPP_DATA_NAME_MAP, releated_name);
for (size_t i = 0; i < data_op_list_.size(); ++i) {
GE_CHECK_NOTNULL(data_op_list_[i]);
if (data_op_list_[i]->GetName() == releated_name) {
GELOGI("Find aipp_data [%s] index %zu from index %u", releated_name.c_str(), i, index);
aipp_index = i;
}
}
if (aipp_index == 0xFFFFFFFF) {
GELOGE(INTERNAL_ERROR, "Can not find aipp data node from index %u", index);
return INTERNAL_ERROR;
}
}
return SUCCESS;
}
void DavinciModel::SetDynamicSize(const std::vector<uint64_t> &batch_num, int32_t dynamic_type) {
batch_size_.clear();
if (batch_num.empty()) {
GELOGD("User has not set dynammic data");
}
for (size_t i = 0; i < batch_num.size(); i++) {
batch_size_.emplace_back(batch_num[i]);
}
dynamic_type_ = dynamic_type;
}
void DavinciModel::GetCurShape(std::vector<int64_t> &batch_info, int32_t &dynamic_type) {
if (batch_size_.empty()) {
GELOGD("User does not set dynamic size");
}
for (size_t i = 0; i < batch_size_.size(); i++) {
GELOGI("Start to get current shape");
batch_info.emplace_back(batch_size_[i]);
}
dynamic_type = dynamic_type_;
}
void DavinciModel::GetModelAttr(std::vector<std::string> &dynamic_output_shape_info) {
for (auto &op : output_op_list_) {
if (op->GetType() != NETOUTPUT) {
continue;
}
GELOGI("Start to get dynamic output dims attr");
if (!AttrUtils::GetListStr(op, ATTR_NAME_DYNAMIC_OUTPUT_DIMS, dynamic_output_shape_info)) {
GELOGD("Can not get dynamic output dims attr");
}
}
}
Status DavinciModel::GetInputOutputDescInfoForZeroCopy(vector<InputOutputDescInfo> &input_desc,
vector<InputOutputDescInfo> &output_desc,
std::vector<uint32_t> &input_formats,
std::vector<uint32_t> &outputFormats) {
if ((data_op_list_.empty()) || (1 != data_op_list_[0]->GetInputsSize())) {
GELOGE(FAILED, "OP List Pointer is null or input_desc size is not 1!");
return FAILED;
}
GE_CHK_STATUS_RET(GetInputDescInfo(input_desc, input_formats), "get input desc info failed");
GE_CHK_STATUS_RET(GetOutputDescInfo(output_desc, outputFormats), "get ouput desc info failed");
GE_CHK_BOOL_RET_STATUS(output_desc.size() == output_memory_size_list_.size(), INTERNAL_ERROR,
"output_desc size[%zu] not equal output_size_list_[%zu] size!", output_desc.size(),
output_memory_size_list_.size());
/// For function zero copy,the momery should be aligned by 512 bytes.
/// And, because of the cce op limit, size should be lager than the real shape size. The memory should be padded by 32
/// bytes.
/// *size equals to ((tensorDesc->dataSize + 2 * 32 - 1) / 32) * 32;
for (size_t i = 0; i < output_memory_size_list_.size(); i++) {
output_desc[i].size = output_memory_size_list_[i];
}
return SUCCESS;
}
void DavinciModel::SetInputDimsInfo(const vector<int64_t> &model_input_dims, Format &format,
InputOutputDescInfo &input) {
uint32_t n, c, h, w;
n = format == FORMAT_NHWC ? NHWC_DIM_N : NCHW_DIM_N;
c = format == FORMAT_NHWC ? NHWC_DIM_C : NCHW_DIM_C;
h = format == FORMAT_NHWC ? NHWC_DIM_H : NCHW_DIM_H;
w = format == FORMAT_NHWC ? NHWC_DIM_W : NCHW_DIM_W;
if (model_input_dims.size() == static_cast<size_t>(NORMAL_TENSOR_SIZE)) {
input.shape_info.num = model_input_dims[n];
input.shape_info.height = model_input_dims[h];
input.shape_info.width = model_input_dims[w];
input.shape_info.channel = model_input_dims[c];
}
for (size_t k = 0; k < model_input_dims.size(); ++k) {
input.shape_info.dims.push_back(model_input_dims[k]);
}
return;
}
void DavinciModel::CreateInputDimsInfo(const OpDescPtr &op_desc, Format format, InputOutputDescInfo &input) {
if (is_new_model_desc_ && op_desc->HasAttr(ATTR_NAME_INPUT_DIMS)) {
// When static aipp is set, need to get the model input dims which processed by aipp
vector<int64_t> model_input_dims;
(void)AttrUtils::GetListInt(op_desc, ATTR_NAME_INPUT_DIMS, model_input_dims);
SetInputDimsInfo(model_input_dims, format, input);
return;
}
// judge if this data is linked dynamic aipp first, multiply batch has been considered
if (op_desc->HasAttr(ATTR_DYNAMIC_AIPP_INPUT_DIMS)) {
vector<int64_t> dynamic_aipp_input_dims;
(void)AttrUtils::GetListInt(op_desc, ATTR_DYNAMIC_AIPP_INPUT_DIMS, dynamic_aipp_input_dims);
SetInputDimsInfo(dynamic_aipp_input_dims, format, input);
return;
} else {
// judge if this data is multiply batch
if (!op_desc->HasAttr(ATTR_MBATCH_ORIGIN_INPUT_DIMS)) {
vector<int64_t> input_dims = op_desc->GetInputDescPtr(0)->GetShape().GetDims();
SetInputDimsInfo(input_dims, format, input);
return;
} else {
vector<int64_t> origin_input_dims;
(void)AttrUtils::GetListInt(op_desc, ATTR_MBATCH_ORIGIN_INPUT_DIMS, origin_input_dims);
SetInputDimsInfo(origin_input_dims, format, input);
return;
}
}
}
Status DavinciModel::GetInputDescInfo(vector<InputOutputDescInfo> &input_desc, std::vector<uint32_t> &formats) {
for (size_t index = 0; index < data_op_list_.size(); ++index) {
InputOutputDescInfo input;
GE_CHECK_NOTNULL(data_op_list_[index]);
GE_CHECK_NOTNULL(data_op_list_[index]->GetInputDescPtr(0));
Format format = data_op_list_[index]->GetInputDescPtr(0)->GetFormat();
CreateInputDimsInfo(data_op_list_[index], format, input);
input.data_type = data_op_list_[index]->GetInputDescPtr(0)->GetDataType();
input.name = data_op_list_[index]->GetName();
int64_t input_size = 0;
GE_CHK_STATUS_RET(TensorUtils::GetSize(*data_op_list_[index]->GetInputDescPtr(0), input_size),
"get input size failed.");
input.size = input_size;
formats.push_back(format);
input_desc.push_back(input);
}
// cause GetInputDescInfo called not only once, set is_new_model_desc_ to false after calc the model input dims
is_new_model_desc_ = false;
return SUCCESS;
}
void DavinciModel::CreateOutput(uint32_t index, OpDescPtr &op_desc, InputOutputDescInfo &output,
uint32_t &format_result) {
/// netoutput input tensor desc
GE_IF_BOOL_EXEC(op_desc->GetInputDescPtr(index) == nullptr, GELOGE(FAILED, "OpDesc GetInputDescPtr is nullptr");
return );
Format format = op_desc->GetInputDescPtr(index)->GetFormat();
GeShape shape = op_desc->GetInputDescPtr(index)->GetShape();
DataType data_type = op_desc->GetInputDescPtr(index)->GetDataType();
int64_t dims[] = {1, 1, 1, 1};
format_result = format;
if (format == FORMAT_ND) { // for ND tensor
for (size_t i = 0; i < shape.GetDimNum() && i < (sizeof(dims) / sizeof(dims[0])); i++) {
dims[i] = shape.GetDim(i);
}
} else { // FOR FORMAT_NHWC or FORMAT_NCHW
dims[0] = shape.GetDim(format == FORMAT_NHWC ? NHWC_DIM_N : NCHW_DIM_N); // 0: first dim
dims[1] = shape.GetDim(format == FORMAT_NHWC ? NHWC_DIM_C : NCHW_DIM_C); // 1: second dim
dims[2] = shape.GetDim(format == FORMAT_NHWC ? NHWC_DIM_H : NCHW_DIM_H); // 2: third dim
dims[3] = shape.GetDim(format == FORMAT_NHWC ? NHWC_DIM_W : NCHW_DIM_W); // 3: forth dim
}
output.shape_info.num = dims[0]; // 0: first dim
output.shape_info.channel = dims[1]; // 1: second dim
output.shape_info.height = dims[2]; // 2: third dim
output.shape_info.width = dims[3]; // 3: forth dim
if (op_desc->GetInputDescPtr(index)->GetFormat() == FORMAT_FRACTAL_Z) { // FraczToHWCK
int64_t k = shape.GetDim(0); // 0: first dim
int64_t c = shape.GetDim(1); // 1: second dim
int64_t h = shape.GetDim(2); // 2: third dim
int64_t w = shape.GetDim(3); // 3: forth dim
output.shape_info.dims.push_back(h);
output.shape_info.dims.push_back(w);
output.shape_info.dims.push_back(c);
output.shape_info.dims.push_back(k);
format_result = FORMAT_HWCN;
} else {
for (size_t j = 0; j < shape.GetDimNum(); j++) {
output.shape_info.dims.push_back(shape.GetDim(j));
}
}
int64_t tensor_size = 0;
(void)TensorUtils::CalcTensorMemSize(shape, format, data_type, tensor_size); // no need to check value
output.size = static_cast<uint64_t>(tensor_size);
output.data_type = op_desc->GetInputDescPtr(index)->GetDataType();
}
Status DavinciModel::GetOutputDescInfo(vector<InputOutputDescInfo> &output_desc, std::vector<uint32_t> &formats) {
GELOGI("Output node size: %zu", output_op_list_.size());
for (size_t i = 0; i < output_op_list_.size(); i++) {
auto &op_desc = output_op_list_[i];
uint32_t out_size = static_cast<uint32_t>(op_desc->GetInputsSize());
for (uint32_t index = 0; index < out_size; index++) {
string output_name;
InputOutputDescInfo output;
uint32_t format_result;
CreateOutput(index, op_desc, output, format_result);
std::vector<std::string> src_name = op_desc->GetSrcName();
std::vector<int64_t> src_index = op_desc->GetSrcIndex();
GE_CHK_BOOL_RET_STATUS(src_name.size() > index && src_index.size() > index, INTERNAL_ERROR,
"construct output_name failed.");
// forward compatbility, if old om has no out_node_name, need to return output follow origin way
if (out_size == out_node_name_.size()) {
// neweast plan, the index will add to name during generate model.
bool contains_colon = out_node_name_[index].find(":") != std::string::npos;
output_name =
contains_colon ? out_node_name_[index] : out_node_name_[index] + ":" + std::to_string(src_index[index]);
} else {
output_name = std::string("output_") + std::to_string(index) + "_" + src_name[index] + "_" +
std::to_string(src_index[index]);
}
output.name = output_name;
output_desc.push_back(output);
formats.push_back(format_result);
}
}
return SUCCESS;
}
ge::Format DavinciModel::GetFormat() {
if ((data_op_list_.empty()) || data_op_list_[0] == nullptr || data_op_list_[0]->GetInputDescPtr(0) == nullptr) {
GELOGW("OP List Pointer is null or input_desc size is not 1!");
return FORMAT_NCHW;
}
return data_op_list_[0]->GetInputDescPtr(0)->GetFormat();
}
Status DavinciModel::CopyInputData(const InputData &input_data, bool device_data) {
rtMemcpyKind_t kind = device_data ? RT_MEMCPY_DEVICE_TO_DEVICE : RT_MEMCPY_HOST_TO_DEVICE;
const std::vector<DataBuffer> &blobs = input_data.blobs;
for (const auto &data : new_input_data_info_) {
if (data.first >= blobs.size()) {
GELOGE(FAILED, "Blobs not match: blobs=%zu, tensor=%zu, index=%u, size=%ld", blobs.size(),
new_input_data_info_.size(), data.first, data.second.GetDataInfo().at(0).first);
return FAILED;
}
const DataBuffer &data_buf = blobs[data.first];
if (data_buf.length == 0) {
GELOGW("No data need to memcpy!");
return SUCCESS;
}
uint64_t data_size = data.second.GetDataSize();
GE_CHK_BOOL_RET_STATUS(data_size >= data_buf.length, PARAM_INVALID,
"input data size(%lu) does not match model required size(%lu), ret failed.", data_buf.length,
data_size);
void *mem_addr = data.second.GetBasicAddr();
void *data_buf_addr = reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(data_buf.data));
uint64_t data_buf_length = data_buf.length;
GELOGI("[IMAS]CopyPlainData memcpy graph_%lu type[F] input[%lu] dst[%p] src[%p] mem_size[%lu] datasize[%lu]",
runtime_param_.graph_id, data.first, mem_addr, data_buf_addr, data_size, data_buf_length);
GE_CHK_RT_RET(rtMemcpy(mem_addr, data_size, data_buf_addr, data_buf_length, kind));
}
return SUCCESS;
}
Status DavinciModel::SyncVarData() {
GELOGI("Sync var data, model id:%u", model_id_);
Status ret = SUCCESS;
OpDescPtr global_step = GetVariableOp(NODE_NAME_GLOBAL_STEP);
if (global_step != nullptr) {
auto v_output_size = ModelUtils::GetOutputSize(global_step);
auto v_output_addr = ModelUtils::GetOutputDataAddrs(runtime_param_, global_step);
if (v_output_size.empty() || v_output_addr.empty()) {
GELOGE(PARAM_INVALID, "global step op:%s not set output", global_step->GetName().c_str());
return PARAM_INVALID;
}
std::vector<uint64_t> v_step;
v_step.push_back(iterator_count_);
GE_CHK_RT_RET(rtMemcpy(v_output_addr[0], v_output_size[0], v_step.data(), v_step.size() * sizeof(uint64_t),
RT_MEMCPY_HOST_TO_DEVICE));
}
for (auto op_desc : variable_op_list_) {
ret =
VarManager::Instance(session_id_)->SyncVarData(runtime_param_.graph_id, op_desc->GetName(), op_desc, mem_base_);
GE_CHK_BOOL_EXEC(ret == SUCCESS, break, "sync var data ret failed, model id:%u, op name:%s.", model_id_,
op_desc->GetName().c_str());
}
return ret;
}
inline int64_t SumSize(const vector<int64_t> &size_list) {
int64_t sum_size = 0;
for (const int64_t &size : size_list) {
sum_size += size;
}
return sum_size;
}
Status DavinciModel::SinkModelProfile() {
// profiling plugin must be registered
Msprof::Engine::Reporter *reporter = PluginImpl::GetPluginReporter();
GE_IF_BOOL_EXEC(reporter == nullptr, GELOGI("Profiling report is nullptr!"); return SUCCESS);
GELOGI("Start collect model load profiling data.");
Msprof::Engine::ReporterData reporter_data{};
// report model data tag name
std::string tag_name;
tag_name.append("model_load_info_").append(std::to_string(this->Id()));
GE_CHK_BOOL_EXEC(memcpy_s(reporter_data.tag, MSPROF_ENGINE_MAX_TAG_LEN, tag_name.c_str(), tag_name.size()) == EOK,
return FAILED, "Sink model tag memcpy error.");
// Model Header
string name;
if (!om_name_.empty()) {
name = om_name_;
} else {
name = name_;
}
size_t name_len = name.size();
reporter_data.deviceId = device_id_;
reporter_data.data = (unsigned char *)&name_len;
reporter_data.dataLen = sizeof(int32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
reporter_data.data = (unsigned char *)name.c_str();
reporter_data.dataLen = name.size();
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
uint32_t model_id = this->Id();
reporter_data.data = (unsigned char *)&model_id;
reporter_data.dataLen = sizeof(uint32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
// Load Start/End Time
int64_t start_time = this->GetLoadBeginTime();
reporter_data.data = (unsigned char *)&start_time;
reporter_data.dataLen = sizeof(int64_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
int64_t end_time = this->GetLoadEndTime();
reporter_data.data = (unsigned char *)&end_time;
reporter_data.dataLen = sizeof(int64_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
int32_t task_num = task_list_.size();
std::multimap<uint32_t, uint32_t> op_id_map;
std::set<uint32_t> task_id_set;
for (int32_t i = 0; i < task_num; i++) {
auto task = task_list_[i];
GE_CHECK_NOTNULL(task);
auto fusion_op_info = task->GetFusionOpInfo();
// when type is RT_MODEL_TASK_KERNEL, ctx is not null
if (fusion_op_info != nullptr) {
uint32_t op_num = fusion_op_info->original_op_names.size();
uint32_t task_id = task->GetTaskID();
if (op_num > 0) {
GELOGI("task.id = %u, opNum = %u", task_id, op_num);
op_id_map.insert(std::make_pair(fusion_op_info->op_index, task_id));
}
}
}
struct memoryInfo {
int64_t input_size;
int64_t output_size;
int64_t weight_size;
int64_t workspace_size;
int64_t total_size;
memoryInfo() : input_size(0), output_size(0), weight_size(0), workspace_size(0), total_size(0) {}
};
using CIT = std::multimap<uint32_t, uint32_t>::const_iterator;
using Range = std::pair<CIT, CIT>;
for (int32_t i = 0; i < task_num; i++) {
auto task = task_list_[i];
GE_CHECK_NOTNULL(task);
auto fusion_op_info = task->GetFusionOpInfo();
if (fusion_op_info != nullptr && fusion_op_info->original_op_names.size() > 0) {
uint32_t task_id = task->GetTaskID();
uint32_t op_num = fusion_op_info->original_op_names.size();
uint32_t task_count = 0;
if (task_id_set.count(task_id) != 0) {
continue;
}
uint32_t op_id = fusion_op_info->op_index;
Range range = op_id_map.equal_range(op_id);
for (CIT range_idx = range.first; range_idx != range.second; ++range_idx) {
task_count++;
uint32_t task_id = range_idx->second;
task_id_set.insert(task_id);
}
// op name after fusion
string fusion_op_name = fusion_op_info->op_name;
int32_t fusion_op_name_len = fusion_op_name.size() == 0 ? 1 : fusion_op_name.size();
reporter_data.data = (unsigned char *)&fusion_op_name_len;
reporter_data.dataLen = sizeof(int32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
reporter_data.data = (unsigned char *)fusion_op_name.c_str();
reporter_data.dataLen = fusion_op_name_len;
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
// original op name before fusion
reporter_data.data = (unsigned char *)&op_num;
reporter_data.dataLen = sizeof(int32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
for (uint32_t k = 0; k < op_num; k++) {
std::string op_name = fusion_op_info->original_op_names[k];
int32_t op_name_len = op_name.size() == 0 ? 1 : op_name.size();
reporter_data.data = (unsigned char *)&op_name_len;
reporter_data.dataLen = sizeof(int32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
reporter_data.data = (unsigned char *)op_name.c_str();
reporter_data.dataLen = op_name_len;
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
}
// stream id info
uint32_t streamId = task->GetStreamId();
reporter_data.data = (unsigned char *)&streamId;
reporter_data.dataLen = sizeof(int32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
// memory info
struct memoryInfo memory_info;
uint32_t op_index = fusion_op_info->op_index;
auto iter = op_list_.find(op_index);
GE_CHK_BOOL_EXEC(iter != op_list_.end(), return FAILED, "index is out of range, index: %u", op_index);
auto op_desc = iter->second;
memory_info.input_size = SumSize(ModelUtils::GetInputSize(op_desc));
memory_info.output_size = SumSize(ModelUtils::GetOutputSize(op_desc));
memory_info.workspace_size = SumSize(ModelUtils::GetWorkspaceSize(op_desc));
memory_info.weight_size = SumSize(ModelUtils::GetWeightSize(op_desc));
memory_info.total_size =
memory_info.weight_size + memory_info.input_size + memory_info.output_size + memory_info.workspace_size;
reporter_data.data = (unsigned char *)&memory_info;
reporter_data.dataLen = sizeof(struct memoryInfo);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
// task info
reporter_data.data = (unsigned char *)&task_count;
reporter_data.dataLen = sizeof(uint32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
Range task_range = op_id_map.equal_range(op_id);
for (CIT idx = task_range.first; idx != task_range.second; ++idx) {
uint32_t task_id = idx->second;
reporter_data.data = (unsigned char *)&task_id;
reporter_data.dataLen = sizeof(uint32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
}
}
}
return SUCCESS;
}
Status DavinciModel::SinkTimeProfile(const InputData &current_data) {
// profiling plugin must be registered
Msprof::Engine::Reporter *reporter = PluginImpl::GetPluginReporter();
GE_IF_BOOL_EXEC(reporter == nullptr, GELOGI("Profiling report is nullptr!"); return SUCCESS);
Msprof::Engine::ReporterData reporter_data{};
// report model data tag name
std::string tag_name;
tag_name.append("model_time_info_")
.append(std::to_string(this->Id()))
.append("_")
.append(std::to_string(current_data.index));
GE_CHK_BOOL_EXEC(memcpy_s(reporter_data.tag, MSPROF_ENGINE_MAX_TAG_LEN, tag_name.c_str(), tag_name.size()) == EOK,
return FAILED, "Sink model tag memcpy error.");
// device id
reporter_data.deviceId = device_id_;
// Model Header
string name;
if (!om_name_.empty()) {
name = om_name_;
} else {
name = name_;
}
size_t name_len = name.size();
reporter_data.data = (unsigned char *)&name_len;
reporter_data.dataLen = sizeof(int32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
reporter_data.data = (unsigned char *)name.c_str();
reporter_data.dataLen = name.size();
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED, "Reporter data fail, model id:%u.",
this->Id());
// request id
uint64_t request_id = current_data.request_id;
reporter_data.data = (unsigned char *)&request_id;
reporter_data.dataLen = sizeof(uint32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED,
"Reporter data fail, model id:%u, data index:%u.", this->Id(), current_data.index);
// thread id
int32_t thread_id = GetDataInputTid();
reporter_data.data = (unsigned char *)&thread_id;
reporter_data.dataLen = sizeof(int32_t);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED,
"Reporter data fail, model id:%u, data index:%u.", this->Id(), current_data.index);
// time info
time_info_.modelId = this->Id();
reporter_data.data = (unsigned char *)&time_info_;
reporter_data.dataLen = sizeof(struct timeInfo);
GE_CHK_BOOL_EXEC(reporter->Report(&reporter_data) == SUCCESS, return FAILED,
"Reporter data fail, model id:%u, data index:%u.", this->Id(), current_data.index);
return SUCCESS;
}
void DavinciModel::SetProfileTime(ModelProcStage stage, int64_t endTime) {
int64_t time = endTime;
if (time == 0) {
mmTimespec timespec = mmGetTickCount();
time = timespec.tv_sec * 1000 * 1000 * 1000 + timespec.tv_nsec; // 1000 ^ 3 converts second to nanosecond
}
switch (stage) {
case MODEL_LOAD_START:
load_begin_time_ = time;
break;
case MODEL_LOAD_END:
load_end_time_ = time;
break;
case MODEL_PRE_PROC_START:
time_info_.processBeginTime = time;
break;
case MODEL_PRE_PROC_END:
time_info_.processEndTime = time;
break;
case MODEL_INFER_START:
time_info_.inferenceBeginTime = time;
break;
case MODEL_INFER_END:
time_info_.inferenceEndTime = time;
break;
case MODEL_AFTER_PROC_START:
time_info_.dumpBeginTime = time;
break;
case MODEL_AFTER_PROC_END:
time_info_.dumpEndTime = time;
break;
default:
break;
}
return;
}
///
/// @ingroup ge
/// @brief send Output Op result to upper layer
/// @already malloced in ModelLoad, no need to malloc again
/// @param [in] data_id: the index of output_data
/// @param [in/out] output_data: real user output_data
/// @param [in] kind: the kind of rtMemcpy
/// @return Status result
/// @author
///
Status DavinciModel::CopyOutputData(uint32_t data_id, OutputData &output_data, rtMemcpyKind_t kind) {
if (output_op_list_.empty()) {
Status ret = SyncVarData();
return ret;
}
output_data.index = data_id;
output_data.model_id = model_id_;
if (output_data.blobs.size() != new_output_data_info_.size()) {
GELOGE(FAILED, "Output data buffer num=%zu not equal model data num=%zu", output_data.blobs.size(),
new_output_data_info_.size());
return FAILED;
}
std::vector<DataBuffer> &blobs = output_data.blobs;
size_t idx = 0;
for (const auto &output : new_output_data_info_) {
if (output.first >= blobs.size()) {
GELOGE(FAILED, "Blobs not match: blobs=%zu, tensor=%zu, index=%u, size=%ld", blobs.size(),
new_input_data_info_.size(), output.first, output.second.GetDataInfo().at(0).first);
return FAILED;
}
if ((kind == RT_MEMCPY_DEVICE_TO_DEVICE) && (copy_only_addrs_.count(output.second.GetBasicAddr()) == 0)) {
continue; // Skip: Feed by zero copy.
}
DataBuffer &buffer = blobs[output.first];
uint64_t mem_size = static_cast<uint64_t>(output.second.GetDataSize());
if ((buffer.length == 0) || (mem_size == 0)) {
GELOGI("Length of data is zero, No need copy. output tensor index=%u", output.first);
continue;
}
if (is_dynamic_) {
GELOGI("No need to check output data size.");
} else if (buffer.length < mem_size) {
GELOGE(FAILED, "Tensor data size=%lu, buffer size=%u", mem_size, buffer.length);
return FAILED;
} else if (buffer.length > mem_size) {
GELOGW("Tensor data size=%lu, buffer size=%u", mem_size, buffer.length);
}
int64_t data_size = output.second.GetDataSize();
if (is_online_infer_dynamic_) {
auto gear_and_real_out_size_info = merge_nodes_gear_and_real_out_size_info_[idx];
data_size = gear_and_real_out_size_info[cur_dynamic_dims_];
}
uint64_t buffer_length = buffer.length;
void *buffer_addr = reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(buffer.data));
GELOGI("[IMAS]CopyPlainData memcpy graph_%u type[F] output[%u] memaddr[%p] mem_size[%ld] datasize[%u]",
runtime_param_.graph_id, output.first, output.second.GetBasicAddr(), data_size, buffer_length);
GE_CHK_RT_RET(rtMemcpy(buffer_addr, buffer_length, output.second.GetBasicAddr(), data_size, kind));
idx++;
}
return SUCCESS;
}
Status DavinciModel::GenOutputTensorInfo(const OpDescPtr &op_desc, uint32_t data_index, OutputData *output_data,
std::vector<ge::OutputTensorInfo> &outputs) {
GE_CHECK_NOTNULL(op_desc);
GE_CHECK_NOTNULL(output_data);
if (output_data->blobs.size() > data_index) {
GELOGI("No need to generate output tensor info, model id:%u", model_id_);
return SUCCESS;
}
std::vector<int64_t> out_buffer_size_vec;
std::vector<std::vector<int64_t>> shape_info_vec;
size_t input_num = op_desc->GetInputsSize();
if (is_getnext_sink_dynamic_) {
input_num = input_num - kGetDynamicDimsCount;
}
for (size_t i = 0; i < input_num; ++i) {
int64_t size = 0;
auto input_desc = op_desc->GetInputDescPtr(i);
GE_CHECK_NOTNULL(input_desc);
auto ret = TensorUtils::GetTensorSizeInBytes(*input_desc, size);
GE_IF_BOOL_EXEC(ret != GRAPH_SUCCESS,
GELOGE(ret, "Get size from TensorDesc failed, op:%s, input id:%zu", op_desc->GetName().c_str(), i);
return ret);
std::vector<int64_t> output_shape = input_desc->GetShape().GetDims();
if (is_online_infer_dynamic_) {
auto gear_and_real_out_size_info = merge_nodes_gear_and_real_out_size_info_[i];
size = gear_and_real_out_size_info[cur_dynamic_dims_];
auto gear_and_real_out_shape_info = merge_nodes_gear_and_real_out_shape_info_[i];
output_shape = gear_and_real_out_shape_info[cur_dynamic_dims_];
is_dynamic_ = true;
}
GELOGI("Output size is %ld, output shape is %s.", size, formats::JoinToString(output_shape).c_str());
out_buffer_size_vec.push_back(size);
shape_info_vec.push_back(output_shape);
}
GELOGI("Output blobs size:%zu, data index:%u, model id:%u", out_buffer_size_vec.size(), data_index, model_id_);
for (size_t i = 0; i < out_buffer_size_vec.size(); ++i) {
std::unique_ptr<uint8_t[]> data_buf(new (std::nothrow) uint8_t[out_buffer_size_vec[i]]);
if (data_buf == nullptr) {
GELOGE(GE_GRAPH_MALLOC_FAILED, "Malloc buffer failed.");
return GE_GRAPH_MALLOC_FAILED;
}
output_data->blobs.push_back({data_buf.get(), static_cast<uint64_t>(out_buffer_size_vec[i]), false});
ge::OutputTensorInfo output;
output.dims = shape_info_vec[i];
output.data = std::move(data_buf);
output.length = out_buffer_size_vec[i];
outputs.emplace_back(std::move(output));
GELOGD("Output index:%zu, output dims is %s, data length:%lu.", i,
formats::JoinToString(output.dims).c_str(), output.length);
}
return SUCCESS;
}
///
/// @ingroup ge
/// @brief send Output Op result to upper layer
/// @already malloced in ModelLoad, no need to malloc again
/// @param [in] data_id: the index of output_data
/// @param [in] rslt_flg: result flag
/// @param [in] seq_end_flag: sequence end flag
/// @param [out] output_data: real user output_data
/// @return Status result
/// @author
///
Status DavinciModel::ReturnResult(uint32_t data_id, const bool rslt_flg, const bool seq_end_flag,
OutputData *output_data) {
GE_CHK_BOOL_EXEC(listener_ != nullptr, return PARAM_INVALID, "listener_ is null.");
std::vector<ge::OutputTensorInfo> outputs;
// return result is not required
if (!rslt_flg && !seq_end_flag) {
GELOGW("Compute failed, model id: %u", model_id_);
auto model_manager = ModelManager::GetInstance();
GE_CHECK_NOTNULL(model_manager);
auto exception_infos = model_manager->GetExceptionInfos();
if (exception_infos.size() > 0) {
GE_CHK_STATUS_RET(data_dumper_.DumpExceptionInfo(exception_infos), "Dump exception info failed");
} else {
GELOGI("Exception info is null");
}
GE_CHK_STATUS(listener_->OnComputeDone(model_id_, data_id, INTERNAL_ERROR, outputs), "OnComputeDone failed.");
return INTERNAL_ERROR;
}
if (output_op_list_.empty()) {
GELOGW("Output tensor list is empty, model id: %u", model_id_);
GE_CHK_STATUS(listener_->OnComputeDone(model_id_, data_id, INTERNAL_ERROR, outputs), "OnComputeDone failed.");
return INTERNAL_ERROR;
}
GE_CHECK_NOTNULL(output_data);
// index of data in output_data
uint32_t data_index = 0;
output_data->index = data_id;
output_data->model_id = model_id_;
is_getnext_sink_dynamic_ = false;
// copy output data from op to designated position
for (auto &op_desc : output_op_list_) {
if (IsGetNextSinkDynamic(op_desc)) {
GELOGD("Reinit cur dynamic dims when getnext sink dynamic.");
is_getnext_sink_dynamic_ = true;
cur_dynamic_dims_.clear();
cur_dynamic_dims_.resize(shape_of_cur_dynamic_dims_);
GE_CHK_RT_RET(rtMemcpy(cur_dynamic_dims_.data(), shape_of_cur_dynamic_dims_ * sizeof(int64_t),
netoutput_last_input_addr_, netoutput_last_input_size_, RT_MEMCPY_DEVICE_TO_HOST));
}
GELOGD("Cur dynamic dims is %s.", formats::JoinToString(cur_dynamic_dims_).c_str());
if (GenOutputTensorInfo(op_desc, data_index, output_data, outputs) != SUCCESS) {
return INTERNAL_ERROR;
}
data_index += op_desc->GetInputsSize();
}
if (CopyOutputData(data_id, *output_data, RT_MEMCPY_DEVICE_TO_HOST) != SUCCESS) {
GE_CHK_STATUS(listener_->OnComputeDone(model_id_, data_id, INTERNAL_ERROR, outputs), "OnComputeDone failed");
return INTERNAL_ERROR;
}
if (seq_end_flag) {
GELOGW("End of sequence, model id: %u", model_id_);
GE_CHK_STATUS(listener_->OnComputeDone(model_id_, data_id, END_OF_SEQUENCE, outputs), "OnCompute Done failed.");
return END_OF_SEQUENCE;
}
GE_CHK_STATUS(listener_->OnComputeDone(model_id_, data_id, SUCCESS, outputs), "OnComputeDone failed");
return SUCCESS;
}
///
/// @ingroup ge
/// @brief return not output to upper layer for cloud case
/// @param [in] data_id
/// @return Status result
///
Status DavinciModel::ReturnNoOutput(uint32_t data_id) {
GELOGI("ReturnNoOutput model id:%u", model_id_);
for (auto op_desc : variable_op_list_) {
Status ret = VarManager::Instance(session_id_)
->SyncBroadCastData2Var(runtime_param_.graph_id, op_desc->GetName(), op_desc, mem_base_);
GE_CHK_BOOL_EXEC(ret == SUCCESS, break, "sync var data ret failed, model id:%u, op name:%s.", model_id_,
op_desc->GetName().c_str());
}
GE_CHK_BOOL_EXEC(listener_ != nullptr, return PARAM_INVALID, "listener_ is null!");
std::vector<ge::OutputTensorInfo> outputs;
GE_CHK_STATUS(listener_->OnComputeDone(model_id_, data_id, SUCCESS, outputs), "OnComputeDone failed.");
return SUCCESS;
}
void *DavinciModel::Run(DavinciModel *model) {
GE_CHK_BOOL_EXEC(model != nullptr,
CsaInteract::GetInstance().WriteErrorCode(FAILED, ERROR_MODULE_FMK, JOBSUBSTATE_GRAPH_EXEC);
return nullptr, "model_pointer is null!")
bool seq_end_flag = false;
uint32_t model_id = model->Id();
uint32_t device_id = model->GetDeviceId();
GELOGI("Model Run thread start, model_id:%u.", model_id);
rtError_t rt_ret = rtSetDevice(static_cast<int32_t>(device_id));
if (rt_ret != RT_ERROR_NONE) {
GELOGE(FAILED, "Model run rtsetdevice failed.");
return nullptr;
}
// DeviceReset before thread run finished!
GE_MAKE_GUARD(not_used_var, [&] { GE_CHK_RT(rtDeviceReset(device_id)); });
while (model->RunFlag()) {
bool rslt_flg = true;
if (model->GetDataInputer() == nullptr) {
GELOGW("Data inputer is nullptr.");
CsaInteract::GetInstance().StoreInternalErrorCode(FAILED, ERROR_MODULE_FMK, JOBSUBSTATE_GRAPH_EXEC);
break;
}
std::shared_ptr<InputDataWrapper> data_wrapper;
Status ret = model->GetDataInputer()->Pop(data_wrapper);
if (data_wrapper == nullptr || ret != SUCCESS) {
GELOGI("data_wrapper is null!");
continue;
}
GELOGI("Getting the input data, model_id:%u", model_id);
GE_IF_BOOL_EXEC(!model->RunFlag(), break);
InputData current_data = data_wrapper->GetInput();
GELOGI("Model thread Run begin, model id:%u, data index:%u.", model_id, current_data.index);
if (model->is_online_infer_dynamic_ && !model->is_getnext_sink_dynamic_) {
model->cur_dynamic_dims_.clear();
GE_IF_BOOL_EXEC(current_data.blobs.empty(), break);
auto shape_data_buffer_data = current_data.blobs.back().data;
auto shape_data_buffer_length = current_data.blobs.back().length;
model->cur_dynamic_dims_.assign(reinterpret_cast<int64_t *>(shape_data_buffer_data),
reinterpret_cast<int64_t *>(shape_data_buffer_data) +
shape_data_buffer_length / sizeof(int64_t));
GELOGD("Data: cur dynamic dims is %s", formats::JoinToString(model->cur_dynamic_dims_).c_str());
}
GE_TIMESTAMP_START(Model_SyncVarData);
ret = model->SyncVarData();
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
ret != SUCCESS, (void)model->ReturnResult(current_data.index, false, false, data_wrapper->GetOutput());
CsaInteract::GetInstance().StoreInternalErrorCode(ret, ERROR_MODULE_FMK, JOBSUBSTATE_GRAPH_EXEC);
continue, "Copy input data to model failed."); // [No need to check value]
GE_IF_BOOL_EXEC(model->is_first_execute_, GE_TIMESTAMP_EVENT_END(Model_SyncVarData, "Model Run SyncVarData"));
GELOGI("Copy input data, model id:%u", model_id);
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(),
model->SetProfileTime(MODEL_PRE_PROC_START));
ret = model->CopyInputData(current_data, false);
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
ret != SUCCESS, (void)model->ReturnResult(current_data.index, false, false, data_wrapper->GetOutput());
CsaInteract::GetInstance().StoreInternalErrorCode(ret, ERROR_MODULE_FMK, JOBSUBSTATE_GRAPH_EXEC);
continue, "Copy input data to model failed."); // [No need to check value]
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), model->SetProfileTime(MODEL_PRE_PROC_END));
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), model->SetProfileTime(MODEL_INFER_START));
if (ProfilingManager::Instance().ProfilingOpTraceOn()) {
GELOGI("GetOpTraceIterNum:%d", ProfilingManager::Instance().GetOpTraceIterNum());
for (int32_t i = 0; i < ProfilingManager::Instance().GetOpTraceIterNum(); i++) {
if (!ProfilingManager::Instance().ProfilingLoadFlag()) {
vector<int32_t> prof_device_id_vec = ProfilingManager::Instance().GetProfilingDeviceId();
for (size_t j = 0; j < prof_device_id_vec.size(); ++j) {
// just profiling, no need to check value
(void)ProfilingManager::Instance().StartProfiling(i, prof_device_id_vec[j]);
}
}
GELOGI("rtModelExecute start.");
rt_ret = rtModelExecute(model->rt_model_handle_, model->rt_model_stream_, 0);
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, rslt_flg = false;
(void)model->ReturnResult(current_data.index, false, false, data_wrapper->GetOutput());
continue); // [No need to check value]
GELOGI("rtModelExecute end");
GELOGI("rtStreamSynchronize start.");
rt_ret = rtStreamSynchronize(model->rt_model_stream_);
if (rt_ret == RT_ERROR_MODEL_ABORT_NORMAL) {
GELOGW("rtStreamSynchronize get result : RT_ERROR_MODEL_ABORT_NORMAL, abort model normal");
} else {
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, rslt_flg = false;
(void)model->ReturnResult(current_data.index, false, seq_end_flag, data_wrapper->GetOutput());
continue); // [No need to check value]
}
GELOGI("rtStreamSynchronize end.");
(void)ProfilingManager::Instance().StopProfiling(); // just profiling, no need to check value
}
} else {
GE_TIMESTAMP_START(rtModelExecute);
GELOGI("rtModelExecute start.");
rt_ret = rtModelExecute(model->rt_model_handle_, model->rt_model_stream_, 0);
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, rslt_flg = false;
(void)model->ReturnResult(current_data.index, false, false, data_wrapper->GetOutput());
CsaInteract::GetInstance().WriteErrorCode(rt_ret, ERROR_MODULE_RUNTIME, JOBSUBSTATE_GRAPH_EXEC);
continue);
GELOGI("rtModelExecute end");
GE_IF_BOOL_EXEC(model->is_first_execute_, GE_TIMESTAMP_EVENT_END(rtModelExecute, "GraphExcute::rtModelExecute"));
GE_TIMESTAMP_START(rtStreamSynchronize);
GELOGI("rtStreamSynchronize start.");
rt_ret = rtStreamSynchronize(model->rt_model_stream_);
if (rt_ret == kEndOfSequence || rt_ret == kEndOfSequenceNew) {
seq_end_flag = true;
}
if (rt_ret == RT_ERROR_MODEL_ABORT_NORMAL) {
GELOGW("rtStreamSynchronize get result : RT_ERROR_MODEL_ABORT_NORMAL, abort model normal");
} else {
GE_IF_BOOL_EXEC(
rt_ret != RT_ERROR_NONE, rslt_flg = false; GELOGI("seq_end_flg: %d", seq_end_flag);
(void)model->ReturnResult(current_data.index, false, seq_end_flag,
data_wrapper->GetOutput()); // [No need to check value]
CsaInteract::GetInstance().StoreInternalErrorCode(rt_ret, ERROR_MODULE_RUNTIME, JOBSUBSTATE_GRAPH_EXEC);
continue);
}
GELOGI("rtStreamSynchronize end.");
GE_IF_BOOL_EXEC(model->is_first_execute_,
GE_TIMESTAMP_EVENT_END(rtStreamSynchronize, "GraphExcute::Wait for rtStreamSynchronize"));
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), model->SetProfileTime(MODEL_INFER_END));
}
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(),
model->SetProfileTime(MODEL_AFTER_PROC_START));
GE_TIMESTAMP_START(ReturnResult3);
// copy output data from device to host
GE_IF_BOOL_EXEC(!model->output_op_list_.empty(),
(void)model->ReturnResult(current_data.index, rslt_flg, false, data_wrapper->GetOutput()))
// copy output data from device to host for variable graph
GE_IF_BOOL_EXEC(model->output_op_list_.empty(), (void)model->ReturnNoOutput(current_data.index));
GE_IF_BOOL_EXEC(model->is_first_execute_,
GE_TIMESTAMP_EVENT_END(ReturnResult3, "GraphExcute::CopyDataFromDeviceToHost"));
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(),
model->SetProfileTime(MODEL_AFTER_PROC_END));
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), (void)model->SinkTimeProfile(current_data));
model->iterator_count_++;
model->is_first_execute_ = false;
GELOGI("run iterator count is %lu", model->iterator_count_);
}
CsaInteract::GetInstance().WriteInternalErrorCode();
GELOGI("Model run end, model id:%u", model->model_id_);
return nullptr;
}
///
/// @ingroup ge
/// @brief call API provided by data inputer to destroy thread
/// @param [in] no
/// @return Status Destroy result
/// @author
///
Status DavinciModel::DestroyThread() {
GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, INTERNAL_ERROR, "data_inputer_ is nullptr.");
run_flg_ = false;
data_inputer_->Stop();
if (thread_id_.joinable()) {
thread_id_.join();
}
return SUCCESS;
}
///
/// @ingroup ge
/// @brief create model std::thread,
/// @brief start to execute Model
/// @param [in] no
/// @return Status create model thread and execute result
/// @author
///
Status DavinciModel::ModelRunStart() {
GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, INTERNAL_ERROR, "data_inputer_ is nullptr.");
LockRunFlg();
GE_MAKE_GUARD(tmp_lock, [&] { UnlockRunFlg(); });
GE_CHK_BOOL_RET_STATUS(!run_flg_, INTERNAL_ERROR, "Model already started.");
run_flg_ = true;
// create stream instance which rt_model_handel is running on
GE_CHK_RT_RET(rtStreamCreate(&rt_model_stream_, priority_));
is_inner_model_stream_ = true;
string opt = "0";
(void)ge::GetContext().GetOption(OPTION_GE_MAX_DUMP_OP_NUM, opt); // option may not be set up, no need to check value
int64_t maxDumpOpNum = std::strtol(opt.c_str(), nullptr, kDecimal);
maxDumpOpNum_ = maxDumpOpNum;
CREATE_STD_THREAD(thread_id_, DavinciModel::Run, this);
GELOGI("model tread create success, model id:%u.", model_id_);
return SUCCESS;
}
///
/// @ingroup ge
/// @brief call API provided by data inputer and destroy model Thread
/// @param [in] no
/// @return Status Destroy result
/// @author
///
Status DavinciModel::ModelRunStop() {
LockRunFlg();
GE_MAKE_GUARD(tmp_lock, [&] { UnlockRunFlg(); });
GE_IF_BOOL_EXEC(!run_flg_, return SUCCESS);
GE_CHK_STATUS_RET(DestroyThread(), "DestoyThead failed.");
return SUCCESS;
}
void DavinciModel::UnbindTaskSinkStream() {
// unbinding hcom stream
UnbindHcomStream();
if (is_stream_list_bind_) {
for (size_t i = 0; i < stream_list_.size(); i++) {
// unbind rt_model_handle and streams
GE_LOGW_IF(rtModelUnbindStream(rt_model_handle_, stream_list_[i]) != RT_ERROR_NONE,
"Unbind stream from model failed! Index: %zu", i);
}
}
if (is_inner_model_stream_) {
if (!input_queue_ids_.empty() || !output_queue_ids_.empty()) {
GE_LOGW_IF(rtModelUnbindStream(rt_model_handle_, rt_model_stream_) != RT_ERROR_NONE, "Unbind stream failed!");
}
// destroy stream that is bound with rt_model
GE_LOGW_IF(rtStreamDestroy(rt_model_stream_) != RT_ERROR_NONE, "Destroy stream for rt_model failed.")
}
if (is_pure_head_stream_ && rt_head_stream_ != nullptr) {
GE_LOGW_IF(rtModelUnbindStream(rt_model_handle_, rt_head_stream_) != RT_ERROR_NONE, "Unbind stream failed!");
GE_LOGW_IF(rtStreamDestroy(rt_head_stream_) != RT_ERROR_NONE, "Destroy stream for rt_model failed.");
rt_head_stream_ = nullptr;
}
if (rt_entry_stream_ != nullptr) {
GE_LOGW_IF(rtModelUnbindStream(rt_model_handle_, rt_entry_stream_) != RT_ERROR_NONE, "Unbind stream failed!");
GE_LOGW_IF(rtStreamDestroy(rt_entry_stream_) != RT_ERROR_NONE, "Destroy stream for rt_model failed.");
rt_entry_stream_ = nullptr;
}
}
Status DavinciModel::CreateKnownZeroCopyMap(const vector<void *> &inputs, const vector<void *> &outputs) {
GELOGI("DavinciModel::CreateKnownZeroCopyMap in.");
if (inputs.size() > data_op_list_.size()) {
GELOGE(FAILED, "input data addr %u should less than input op number %u.", inputs.size(), data_op_list_.size());
return FAILED;
}
// remove zero copy addr in last iteration
knonw_input_data_info_.clear();
knonw_output_data_info_.clear();
for (size_t i = 0; i < inputs.size(); ++i) {
const vector<void *> addr_list = ModelUtils::GetOutputDataAddrs(runtime_param_, data_op_list_[i]);
knonw_input_data_info_[addr_list[kDataIndex]] = inputs[i];
GELOGI("DavinciModel::CreateKnownZeroCopyMap input %d,v addr %p,p addr %p .", i, addr_list[kDataIndex], inputs[i]);
}
if (output_op_list_.size() < kOutputNum) {
GELOGW("output op num in graph is %u.", output_op_list_.size());
return SUCCESS;
}
const vector<void *> addr_list = ModelUtils::GetInputDataAddrs(runtime_param_, output_op_list_[kDataIndex]);
for (size_t i = 0; i < addr_list.size() && i < outputs.size(); ++i) {
knonw_output_data_info_[addr_list[i]] = outputs[i];
GELOGI("DavinciModel::CreateKnownZeroCopyMap output %d,v addr %p,p addr %p .", i, addr_list[i], outputs[i]);
}
GELOGI("DavinciModel::CreateKnownZeroCopyMap success.");
return SUCCESS;
}
Status DavinciModel::UpdateKnownZeroCopyAddr() {
for (size_t i = 0; i < total_io_addrs_.size(); ++i) {
auto it_in = knonw_input_data_info_.find(total_io_addrs_[i]);
if (it_in != knonw_input_data_info_.end()) {
GELOGI("DavinciModel::UpdateKnownZeroCopyAddr input %d,v addr %p,p addr %p .", i, total_io_addrs_[i],
knonw_input_data_info_.at(total_io_addrs_[i]));
total_io_addrs_[i] = knonw_input_data_info_.at(total_io_addrs_[i]);
}
auto it_out = knonw_output_data_info_.find(total_io_addrs_[i]);
if (it_out != knonw_output_data_info_.end()) {
GELOGI("DavinciModel::UpdateKnownZeroCopyAddr output %d,v addr %p,p addr %p .", i, total_io_addrs_[i],
knonw_output_data_info_.at(total_io_addrs_[i]));
total_io_addrs_[i] = knonw_output_data_info_.at(total_io_addrs_[i]);
}
}
GELOGI("DavinciModel::UpdateKnownZeroCopyAddr success.");
return SUCCESS;
}
Status DavinciModel::UpdateKnownNodeArgs(const vector<void *> &inputs, const vector<void *> &outputs) {
GELOGI("DavinciModel::UpdateKnownNodeArgs in");
GE_CHK_STATUS_RET(CreateKnownZeroCopyMap(inputs, outputs),
"DavinciModel::UpdateKnownNodeArgs create map for input/output zero copy.");
if (!base_addr_not_changed_) {
total_io_addrs_.clear();
orig_total_io_addrs_.clear();
for (size_t task_index = 0; task_index < task_list_.size(); ++task_index) {
auto &task = task_list_[task_index];
if (task != nullptr) {
Status ret = task->UpdateArgs();
if (ret != SUCCESS) {
GELOGE(FAILED, "task %d created by davinci model is nullptr.", task_index);
return FAILED;
}
}
}
// cache latest iterator io addr
orig_total_io_addrs_ = total_io_addrs_;
} else {
total_io_addrs_ = orig_total_io_addrs_;
}
GE_CHK_STATUS_RET(UpdateKnownZeroCopyAddr(), "DavinciModel::UpdateKnownZeroCopyAddr failed.");
if (total_args_size_ == 0) {
GELOGW("DavinciModel::UpdateKnownNodeArgs device args %p, dst size %u, pass rtMemcpy.", args_, total_args_size_);
} else {
uint32_t total_addr_size = total_io_addrs_.size() * sizeof(uint64_t);
GELOGI("DavinciModel::UpdateKnownNodeArgs device args %p, dst size %u, src size %u", args_, total_args_size_,
total_addr_size);
Status rt_ret =
rtMemcpy(args_, total_args_size_, total_io_addrs_.data(), total_addr_size, RT_MEMCPY_HOST_TO_DEVICE);
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(rt_ret, "rtMemcpy error, ret: Ox%X", rt_ret); return FAILED;)
}
GELOGI("DavinciModel::UpdateKnownNodeArgs success");
return SUCCESS;
}
Status DavinciModel::InitTaskInfo(domi::ModelTaskDef &model_task_def) {
GELOGI("InitTaskInfo in, task size %zu", model_task_def.task().size());
task_list_.resize(model_task_def.task_size());
for (int i = 0; i < model_task_def.task_size(); ++i) {
// dynamic shape will create task_list_ before
const domi::TaskDef &task = model_task_def.task(i);
if (this->task_list_[i] == nullptr) {
task_list_[i] = TaskInfoFactory::Instance().Create(static_cast<rtModelTaskType_t>(task.type()));
}
GE_CHECK_NOTNULL(task_list_[i]);
Status ret = task_list_[i]->Init(task, this);
if (ret != SUCCESS) {
GELOGE(ret, "Task index %d init failed.", i);
return ret;
}
}
GELOGI("InitTaskInfo out");
return SUCCESS;
}
Status DavinciModel::MallocKnownArgs() {
GELOGI("DavinciModel::MallocKnownArgs in");
const auto &model_task_def = ge_model_->GetModelTaskDefPtr();
if (model_task_def->task_size() == 0) {
GELOGW("DavinciModel::MallocKnownArgs davincimodel has no task info.");
return SUCCESS;
}
task_list_.resize(model_task_def->task_size());
for (int32_t i = 0; i < model_task_def->task_size(); ++i) {
const domi::TaskDef &taskdef = model_task_def->task(i);
task_list_[i] = TaskInfoFactory::Instance().Create(static_cast<rtModelTaskType_t>(taskdef.type()));
GE_CHECK_NOTNULL(task_list_[i]);
Status ret = task_list_[i]->CalculateArgs(taskdef, this);
if (ret != SUCCESS) {
GELOGE(ret, "TaskInfo CalculateArgs failed.");
return ret;
}
}
// malloc args memory
if (total_args_size_ == 0) {
GELOGW("DavinciModel::MallocKnownArgs total_args_size_ equals to zero.");
return SUCCESS;
}
rtError_t rt_ret = rtMalloc(&args_, total_args_size_, RT_MEMORY_HBM);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
// malloc fixed addr memory, eg: rts op
if (total_fixed_addr_size_ != 0) {
GELOGI("Begin to allocate fixed addr.");
rt_ret = rtMalloc(&fixed_addrs_, total_fixed_addr_size_, RT_MEMORY_HBM);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
}
GELOGI("DavinciModel::MallocKnownArgs success, total args size %u. total fixed addr size %ld", total_args_size_,
total_fixed_addr_size_);
return SUCCESS;
}
Status DavinciModel::DistributeTask() {
GELOGI("do Distribute.");
for (auto &task : cpu_task_list_) {
if (task == nullptr) {
GELOGW("task is null");
continue;
}
GE_CHK_STATUS_RET(task->Distribute());
}
task_desc_info_.clear();
bool flag = GetL1FusionEnableOption();
char *skt_enable_env = std::getenv("SKT_ENABLE");
int64_t env_flag = (skt_enable_env != nullptr) ? std::strtol(skt_enable_env, nullptr, kDecimal) : 0;
if (env_flag != 0) {
flag = true;
}
const auto &model_task_def = ge_model_->GetModelTaskDefPtr();
GELOGI("there are %zu task need to save.", task_list_.size());
for (size_t task_index = 0; task_index < task_list_.size(); ++task_index) {
auto &task = task_list_.at(task_index);
GE_CHK_STATUS_RET(task->Distribute(), "Task[%zu] distribute fail", task_index);
// for data dump
auto op_index = std::max(model_task_def->task(task_index).kernel().context().op_index(),
model_task_def->task(task_index).kernel_ex().op_index());
OpDescPtr op = GetOpByIndex(op_index);
GE_CHECK_NOTNULL(op);
SaveDumpOpInfo(runtime_param_, op, task->GetTaskID(), task->GetStreamId());
if (reinterpret_cast<void *>(task->GetDumpArgs()) != nullptr) {
bool call_dump = GetDumpProperties().IsLayerNeedDump(name_, om_name_, op->GetName()) && task->CallSaveDumpInfo();
if (call_dump || is_op_debug_reg_) {
SaveDumpTask(task->GetTaskID(), task->GetStreamId(), op, task->GetDumpArgs());
}
}
// Load task info for profiling
TaskDescInfo task_desc_info;
if (!om_name_.empty()) {
task_desc_info.model_name = om_name_;
} else {
task_desc_info.model_name = name_;
}
task_desc_info.op_name = op->GetName();
task_desc_info.block_dim = model_task_def->task(task_index).kernel().block_dim();
task_desc_info.task_id = task->GetTaskID();
task_desc_info.stream_id = task->GetStreamId();
task_desc_info_.emplace_back(task_desc_info);
if (flag) {
if (task->GetSktTaskID() != 0xFFFFFFFF) {
TaskDescInfo task_desc_info;
string op_name = "super_kernel_" + to_string(task_index);
task_desc_info.op_name = op_name;
task_desc_info.task_id = task->GetSktTaskID();
task_desc_info_.emplace_back(task_desc_info);
}
}
}
// launch dump kernel to aicpu
GE_CHK_STATUS_RET(data_dumper_.LoadDumpInfo(), "Load dump info failed.");
return SUCCESS;
}
void DavinciModel::SetEndGraphId(uint32_t task_id, uint32_t stream_id) {
auto all_dump_model = GetDumpProperties().GetAllDumpModel();
bool findByOmName = all_dump_model.find(om_name_) != all_dump_model.end();
bool findByModelName = all_dump_model.find(name_) != all_dump_model.end();
if (all_dump_model.find(ge::DUMP_ALL_MODEL) != all_dump_model.end() || findByOmName || findByModelName) {
GELOGI("start save end_graph_info to dumper, task_id is %u, stream_id is %u", task_id, stream_id);
data_dumper_.SaveEndGraphId(task_id, stream_id);
}
}
///
/// @ingroup ge
/// @brief Set copy only for No task feed NetOutput address.
/// @return None.
///
void DavinciModel::SetCopyOnlyOutput() {
for (const auto &output_outside_addrs : new_output_outside_addrs_) {
ZeroCopyOffset output_outside = output_outside_addrs.second;
for (uint32_t out_count = 0; out_count < output_outside.GetAddrCount(); ++out_count) {
auto &addrs_mapping_list = output_outside.GetOutsideAddrs();
std::map<const void *, std::vector<void *>> virtual_args_addrs = addrs_mapping_list[out_count];
for (const auto &virtual_args_addr : virtual_args_addrs) {
const auto &args_addrs = virtual_args_addr.second;
if (args_addrs.empty()) { // No task feed Output addr, Need copy directly.
GELOGI("[ZCPY] just copy %p to netoutput.", virtual_args_addr.first);
copy_only_addrs_.insert(virtual_args_addr.first);
}
}
}
}
}
///
/// @ingroup ge
/// @brief Set disabled input zero copy addr.
/// @param [in] const void *addr: address of task
/// @return None.
///
void DavinciModel::DisableZeroCopy(const void *addr) {
if (real_virtual_addrs_.find(addr) == real_virtual_addrs_.end()) {
return;
}
// Data link to RTS Op directly.
std::lock_guard<std::mutex> lock(outside_addrs_mutex_);
GELOGI("[ZCPY] disable zero copy of %p.", addr);
copy_only_addrs_.insert(addr);
}
///
/// @ingroup ge
/// @brief Save outside address used info for ZeroCopy.
/// @param [in] const OpDescPtr &op_desc: current op desc
/// @param [in] const std::vector<void *> &outside_addrs: address of task
/// @param [in] const void *info: task args
/// @param [in] const char *args: task args
/// @param [in] size_t size: size of task args
/// @param [in] size_t offset: offset of task args
/// @return None.
///
void DavinciModel::SetZeroCopyAddr(const OpDescPtr &op_desc, const std::vector<void *> &outside_addrs, const void *info,
void *args, size_t size, size_t offset) {
// Internal call has ensured that op_desc is not nullptr
GELOGD("[ZCPY] SetZeroCopyAddr for %s.", op_desc->GetName().c_str());
size_t nums = outside_addrs.size();
ZeroCopyTask zero_copy_task(op_desc->GetName(), static_cast<uint8_t *>(args), size);
for (size_t i = 0; i < nums; ++i) {
std::lock_guard<std::mutex> lock(outside_addrs_mutex_);
for (auto &input_outside_addrs : new_input_outside_addrs_) {
ZeroCopyOffset &input_outside = input_outside_addrs.second;
bool ret = input_outside.SetOutsideAddrsValue(zero_copy_task, outside_addrs[i], args, offset + i * kAddrLen);
if (ret) {
void *args_val = static_cast<uint8_t *>(args) + offset + i * kAddrLen;
SetBatchLabelAddr(op_desc, reinterpret_cast<uintptr_t>(args_val));
}
}
for (auto &output_outside_addrs : new_output_outside_addrs_) {
ZeroCopyOffset &output_outside = output_outside_addrs.second;
bool ret = output_outside.SetOutsideAddrsValue(zero_copy_task, outside_addrs[i], args, offset + i * kAddrLen);
if (ret) {
void *args_val = static_cast<uint8_t *>(args) + offset + i * kAddrLen;
SetBatchLabelAddr(op_desc, reinterpret_cast<uintptr_t>(args_val));
}
}
}
auto it = zero_copy_op_id_batch_label_.find(op_desc->GetId());
if (it == zero_copy_op_id_batch_label_.end()) {
zero_copy_task.SetBatchLabel(kDefaultBatchLable);
} else {
zero_copy_task.SetBatchLabel(it->second);
}
std::lock_guard<std::mutex> lock(outside_addrs_mutex_);
if (zero_copy_task.IsTaskArgsSet()) {
zero_copy_task.SetOriginalArgs(info, offset + nums * kAddrLen);
zero_copy_tasks_.emplace_back(zero_copy_task);
}
}
void DavinciModel::SetBatchLabelAddr(const OpDescPtr &op_desc, uintptr_t addr) {
// Establish a mapping between batch label and zero copy address for multi-batch scenes
auto it = zero_copy_op_id_batch_label_.find(op_desc->GetId());
if (it == zero_copy_op_id_batch_label_.end()) {
return;
}
const string &batch_label = it->second;
auto iter = zero_copy_batch_label_addrs_.find(batch_label);
if (iter != zero_copy_batch_label_addrs_.end()) {
iter->second.insert(addr);
GELOGD("[ZCPY] Set zero copy batch label and addrs success, batch label: %s, op name:%s.", batch_label.c_str(),
op_desc->GetName().c_str());
} else {
set<uintptr_t> addrs = {addr};
zero_copy_batch_label_addrs_.emplace(pair<string, set<uintptr_t>>(batch_label, addrs));
GELOGD("[ZCPY] New added zero copy batch label and addrs success, batch label: %s, op name:%s.",
batch_label.c_str(), op_desc->GetName().c_str());
}
}
///
/// @ingroup ge
/// @brief Copy Check input size and model op size.
/// @param [in] const int64_t &input_size: input size.
/// @param [in] const int64_t &op_size: model op size.
/// @param [in] is_dynamic: dynamic batch input flag.
/// @return true if success
///
bool DavinciModel::CheckInputAndModelSize(const int64_t &input_size, const int64_t &op_size, bool is_dynamic) {
if (is_dynamic) { // dynamic is max size.
GELOGI("No need to check input and model size.");
return true;
}
if (input_size > op_size) {
GELOGW(
"Input size [%u] is bigger than om size need [%u], "
"MAY cause inference result ERROR, please check model input",
input_size, op_size);
}
bool is_dynamic_aipp = false;
for (const auto &op_desc : data_op_list_) {
if (op_desc->GetType() == AIPP_DATA_TYPE) {
GELOGI("This is dynamic aipp model.");
is_dynamic_aipp = true;
break;
}
}
if (is_dynamic_aipp) {
GELOGI("This is dynamic aipp model, no need to judge smaller input size");
return true;
}
// Judge overflow first
if (input_size > (INT64_MAX - kDataMemAlignSizeCompare)) {
GELOGI("The Input size [%ld] is smaller than model size [%ld] and is in the range of 64 bytes", input_size,
op_size);
return true;
}
// The input and model input size can not be exactly equal because user input is not definite.
if ((input_size + kDataMemAlignSizeCompare) < op_size) {
GELOGE(FAILED, "Input size [%ld] can not be smaller than op size [%ld] after 64-byte alignment", input_size,
op_size);
return false;
}
return true;
}
///
/// @ingroup ge
/// @brief Copy Inputs and Outputs addr to model for direct use.
/// @param [in] const InputData &input_data: model input data.
/// @param [in] OutputData &output_data: model output data.
/// @param [in] bool is_dynamic_input: whether is dynamic input, true: is dynamic input; false: not is dynamic input
/// @return SUCCESS handle successfully / PARAM_INVALID for failed
///
Status DavinciModel::CopyModelData(const InputData &input_data, OutputData &output_data, bool is_dynamic) {
if (UpdateIoTaskArgs(new_input_data_info_, true, input_data.blobs, is_dynamic, input_data.batch_label) != SUCCESS) {
GELOGE(PARAM_INVALID, "[ZCPY] Update input data to model failed.");
return PARAM_INVALID;
}
if (UpdateIoTaskArgs(new_output_data_info_, false, output_data.blobs, is_dynamic, input_data.batch_label) !=
SUCCESS) {
GELOGE(PARAM_INVALID, "[ZCPY] Update output data to model failed.");
return PARAM_INVALID;
}
for (ZeroCopyTask &task : zero_copy_tasks_) {
GE_CHK_STATUS_RET(task.DistributeParam(is_async_mode_, rt_model_stream_), "[ZCPY] Update args failed.");
}
output_data.index = input_data.index;
output_data.model_id = model_id_;
return SUCCESS;
}
///
/// @ingroup ge
/// @brief Copy Data addr to model for direct use.
/// @param [in] data_info: model memory addr/size map { data_index, { tensor_size, tensor_addr } }.
/// @param [in] is_input: input data or output data
/// @param [in] blobs: user input/output data list.
/// @param [in] is_dynamic: whether is dynamic input, true: is dynamic input; false: not is dynamic input
/// @param [in] batch_label: batch label for multi-batch scenes
/// @return SUCCESS handle successfully / others handle failed
///
Status DavinciModel::UpdateIoTaskArgs(const std::map<uint32_t, ZeroCopyOffset> &data_info, bool is_input,
const vector<DataBuffer> &blobs, bool is_dynamic, const string &batch_label) {
string input_or_output = "input";
is_input ? input_or_output = "input" : input_or_output = "output";
if (blobs.size() != data_info.size()) {
GELOGE(FAILED, "Verify %s data num failed: model requires %zu, but user actually feeds %zu",
input_or_output.c_str(), data_info.size(), blobs.size());
return FAILED;
}
for (const auto &data : data_info) {
if (data.first >= blobs.size()) { // check data index.
GELOGE(FAILED, "Verify %s data num failed: can not find No.%zu data, because user only feeds %zu",
input_or_output.c_str(), data.first, blobs.size());
return FAILED;
}
const DataBuffer &buffer = blobs[data.first]; // index of data.
if (buffer.data == nullptr) {
GELOGE(FAILED, "data_buf.data is nullptr, index=%u", data.first);
return FAILED;
}
if (!CheckInputAndModelSize(buffer.length, data.second.GetDataSize(), is_dynamic)) {
GELOGE(FAILED, "Check input size and model size failed");
return FAILED;
}
void *basic_addr = data.second.GetBasicAddr();
uint64_t data_size = data.second.GetDataSize();
if (copy_only_addrs_.count(basic_addr) > 0) {
if (is_input) {
GELOGI("[IMAS] Find addr %p need direct copy from user malloc input %p", basic_addr, buffer.data);
if (rtMemcpy(basic_addr, data_size, buffer.data, buffer.length, RT_MEMCPY_DEVICE_TO_DEVICE) != RT_ERROR_NONE) {
GELOGE(FAILED, "Non-zero copy data node copy failed");
return FAILED;
}
}
GELOGI("No need to exeucte zero copy task because this addr %p need direct copy.", basic_addr);
continue;
}
for (size_t count = 0; count < data.second.GetDataCount(); ++count) {
int64_t size = data.second.GetDataInfo().at(count).first;
void *addr = data.second.GetDataInfo().at(count).second;
void *buffer_addr = reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(buffer.data) +
data.second.GetRelativeOffset().at(count));
GELOGI("[ZCPY] Copy %s blobs_index %u, virtual_addr: %p, size: %ld, user_data_addr: %p", input_or_output.c_str(),
data.first, addr, size, buffer_addr);
// For input data, just copy for rts task.
for (ZeroCopyTask &task : zero_copy_tasks_) {
if (task.GetBatchLabel() != kDefaultBatchLable && task.GetBatchLabel() != batch_label) {
continue;
}
uintptr_t addr_val = reinterpret_cast<uintptr_t>(addr);
if (task.UpdateTaskParam(addr_val, buffer_addr, zero_copy_batch_label_addrs_, batch_label) != SUCCESS) {
return FAILED;
}
}
}
}
return SUCCESS;
}
///
/// @ingroup ge
/// @brief get unique identification for op when load two or more models
/// @param [in] const OpDescPtr: current op.
/// @param [in] string identification: unique identification for current op.
/// @return SUCCESS handle successfully / others handle failed
///
void DavinciModel::GetUniqueId(const OpDescPtr &op_desc, std::string &unique_identification) {
std::string session_graph_id;
GE_IF_BOOL_EXEC(AttrUtils::GetStr(*op_desc, ATTR_NAME_SESSION_GRAPH_ID, session_graph_id),
GELOGD("Get original type of session_graph_id."));
if (session_graph_id.empty()) {
return;
} else if (session_graph_id.find("-1") != string::npos) {
unique_identification = session_graph_id + "_" + to_string(model_id_);
} else {
unique_identification = session_graph_id;
}
}
///
/// @ingroup ge
/// @brief For TVM Op, avoid Addr Reuse.
/// @return void*
///
const char *DavinciModel::GetRegisterStub(const string &binfile, const string &session_graph_id) {
string binfile_key;
if (session_graph_id.empty()) {
binfile_key = binfile;
} else {
binfile_key = session_graph_id + "_" + binfile;
}
auto it = tvm_bin_kernel_.find(binfile_key);
if (it != tvm_bin_kernel_.end()) {
return it->c_str();
} else {
it = tvm_bin_kernel_.insert(tvm_bin_kernel_.end(), binfile_key);
return it->c_str();
}
}
///
/// @ingroup ge
/// @brief Constant Op Init.
/// @return Status
///
Status DavinciModel::InitConstant(const OpDescPtr &op_desc) {
auto v_weights = ModelUtils::GetWeights(op_desc);
auto v_output_size = ModelUtils::GetOutputSize(op_desc);
auto v_output_addr = ModelUtils::GetOutputDataAddrs(runtime_param_, op_desc);
GE_IF_BOOL_EXEC(v_weights.empty() || v_output_size.empty() || v_output_addr.empty(),
GELOGE(PARAM_INVALID, "const op:%s not set output", op_desc->GetName().c_str());
return PARAM_INVALID;);
GeTensor *tensor = const_cast<GeTensor *>(v_weights[0].get());
GE_IF_BOOL_EXEC(static_cast<size_t>(v_output_size[0]) < tensor->GetData().size(),
GELOGE(PARAM_INVALID, "output size:%u less than weight data size:%zu", v_output_size[0],
tensor->GetData().size());
return PARAM_INVALID;);
GE_IF_BOOL_EXEC(tensor->GetData().size() == 0, GELOGW("const op:%s has no weight data.", op_desc->GetName().c_str());
return SUCCESS;);
auto desc = tensor->GetTensorDesc();
if (desc.GetDataType() == DT_STRING) {
GeShape tensor_shape = desc.GetShape();
/// if tensor is a scaler, it's shape size if zero, according ge_tensor.cc.
/// the logic of GetShapeSize is wrong, the scaler tensor's GetShapeSize is zero
/// and that of unknown shape is zero too.
/// unknown shape will not appear here, so we can use zero judge a tensor is scaler or not
int64_t elem_num = tensor_shape.GetShapeSize();
if (elem_num == 0 && tensor_shape.GetDims().size() == 0) {
elem_num = 1;
}
uint64_t *buff = reinterpret_cast<uint64_t *>(tensor->MutableData().data());
GE_CHK_BOOL_RET_STATUS(ge::CheckInt64Uint32MulOverflow(elem_num, kBytes) == SUCCESS, FAILED,
"Shape size is invalid");
uint64_t offset = static_cast<uint64_t>(elem_num * kBytes);
uint64_t hbm_raw_data_base_addr =
reinterpret_cast<uint64_t>(reinterpret_cast<uintptr_t>(v_output_addr[0])) + offset;
for (int64_t i = elem_num - 1; i >= 0; --i) {
buff[i] = hbm_raw_data_base_addr + (buff[i] - buff[0]);
}
}
GELOGI("[IMAS]InitConstant memcpy graph_%u type[V] name[%s] output[%d] memaddr[%p] mem_size[%u] datasize[%zu]",
runtime_param_.graph_id, op_desc->GetName().c_str(), 0, v_output_addr[0], v_output_size[0],
tensor->GetData().size());
GE_CHK_RT_RET(rtMemcpy(v_output_addr[0], v_output_size[0], tensor->GetData().data(), tensor->GetData().size(),
RT_MEMCPY_HOST_TO_DEVICE));
return SUCCESS;
}
///
/// @ingroup ge
/// @brief TVM Op Init.
/// @return Status
///
Status DavinciModel::InitTbeHandle(const OpDescPtr &op_desc) {
auto kernel = ge_model_->GetTBEKernelStore().FindKernel(op_desc->GetName());
auto tbe_kernel = (kernel != nullptr) ? kernel : op_desc->TryGetExtAttr(OP_EXTATTR_NAME_TBE_KERNEL, TBEKernelPtr());
if (tbe_kernel == nullptr) {
GELOGE(INTERNAL_ERROR, "TBE: %s can't find tvm bin file!", op_desc->GetName().c_str());
return INTERNAL_ERROR;
}
std::string session_graph_model_id;
GetUniqueId(op_desc, session_graph_model_id);
const char *bin_file_key = GetRegisterStub(op_desc->GetName(), session_graph_model_id); // from set, always valid.
TBEHandleStore &kernel_store = TBEHandleStore::GetInstance();
std::lock_guard<std::mutex> lock(tvm_bin_mutex_);
if (rtQueryFunctionRegistered(bin_file_key) != RT_ERROR_NONE) {
void *bin_handle = nullptr;
if (!kernel_store.FindTBEHandle(bin_file_key, bin_handle)) {
GELOGI("TBE: can't find the kernel_name[%s] in HandleMap", bin_file_key);
rtDevBinary_t binary;
std::string json_string;
GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc, TVM_ATTR_NAME_MAGIC, json_string),
GELOGI("Get original type of session_graph_id."));
if (json_string == "RT_DEV_BINARY_MAGIC_ELF_AICPU") {
binary.magic = RT_DEV_BINARY_MAGIC_ELF_AICPU;
} else if (json_string == "RT_DEV_BINARY_MAGIC_ELF") {
binary.magic = RT_DEV_BINARY_MAGIC_ELF;
} else if (json_string == "RT_DEV_BINARY_MAGIC_ELF_AIVEC") {
binary.magic = RT_DEV_BINARY_MAGIC_ELF_AIVEC;
} else {
GELOGE(PARAM_INVALID, "TBE: Invalid parameter magic number! json: %s", json_string.c_str());
return PARAM_INVALID;
}
binary.version = 0;
binary.data = tbe_kernel->GetBinData();
binary.length = tbe_kernel->GetBinDataSize();
GELOGI("TBE: binary.length: %lu", binary.length);
GE_CHK_RT_RET(rtDevBinaryRegister(&binary, &bin_handle));
std::string meta_data;
GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc, TVM_ATTR_NAME_METADATA, meta_data),
GELOGI("Get original type of json_string"));
GELOGI("TBE: meta data: %s", meta_data.empty() ? "null" : meta_data.c_str());
GE_IF_BOOL_EXEC(!meta_data.empty(), GE_CHK_RT_RET(rtMetadataRegister(bin_handle, meta_data.c_str())));
kernel_store.StoreTBEHandle(bin_file_key, bin_handle, tbe_kernel);
} else {
GELOGI("TBE: find the kernel_name[%s] in HandleMap", bin_file_key);
kernel_store.ReferTBEHandle(bin_file_key);
}
std::string kernel_name;
GE_IF_BOOL_EXEC(AttrUtils::GetStr(op_desc, op_desc->GetName() + "_kernelname", kernel_name),
GELOGI("Get original type of kernel_name"));
GELOGI("TBE: binfile_key=%s, kernel_name=%s", bin_file_key, kernel_name.c_str());
GE_CHK_RT_RET(rtFunctionRegister(bin_handle, bin_file_key, bin_file_key, kernel_name.c_str(), 0));
used_tbe_handle_map_[bin_file_key] = 1; // Init used num to 1.
return SUCCESS;
}
// Kernel registed, Increase used num in store.
StoreTbeHandle(bin_file_key);
return SUCCESS;
}
void DavinciModel::StoreTbeHandle(const std::string &handle_key) {
// Online mode FE may call rtFunctionRegister.
TBEHandleStore &kernel_store = TBEHandleStore::GetInstance();
auto it = used_tbe_handle_map_.find(handle_key);
if (it != used_tbe_handle_map_.end()) {
// GE registered, increase reference.
kernel_store.ReferTBEHandle(handle_key);
it->second++;
return;
}
void *bin_handle = nullptr;
if (kernel_store.FindTBEHandle(handle_key, bin_handle)) {
// GE registered, increase reference.
used_tbe_handle_map_[handle_key] = 1; // Init used num to 1.
kernel_store.ReferTBEHandle(handle_key);
}
}
void DavinciModel::CleanTbeHandle() {
TBEHandleStore &kernel_store = TBEHandleStore::GetInstance();
kernel_store.EraseTBEHandle(used_tbe_handle_map_);
used_tbe_handle_map_.clear();
tvm_bin_kernel_.clear();
}
///
/// @ingroup ge
/// @brief insert active_stream_indication_
/// @return Status
///
Status DavinciModel::InitStreamActive(const OpDescPtr &op_desc) {
if (op_desc->HasAttr(ATTR_NAME_SWITCH_BRANCH_NODE_LABEL)) {
std::vector<uint32_t> active_stream_list;
GE_CHK_BOOL_EXEC(AttrUtils::GetListInt(op_desc, ATTR_NAME_ACTIVE_STREAM_LIST, active_stream_list),
return INTERNAL_ERROR, "StreamActiveOp get attr ACTIVE_STREAM failed.");
for (size_t j = 0; j < active_stream_list.size(); ++j) {
active_stream_indication_.insert(active_stream_list[j]);
GELOGI("flowctrl_op_index_map node:%s, active_stream_id=%u.", op_desc->GetName().c_str(), active_stream_list[j]);
}
}
return SUCCESS;
}
Status DavinciModel::InitStreamSwitch(const OpDescPtr &op_desc) {
std::vector<uint32_t> active_stream_list;
GE_LOGI_IF(!ge::AttrUtils::GetListInt(op_desc, ATTR_NAME_ACTIVE_STREAM_LIST, active_stream_list),
"GetInt ACTIVE_STREAM_LIST failed.");
if (active_stream_list.size() != kTrueBranchStreamNum) {
GELOGE(INTERNAL_ERROR, "Stream num of switch true branch must be %u.", kTrueBranchStreamNum);
return INTERNAL_ERROR;
}
uint32_t true_stream_id = active_stream_list.front();
active_stream_indication_.insert(true_stream_id);
GELOGI("flowctrl_op_index_map node:%s, true_stream_id=%u.", op_desc->GetName().c_str(), true_stream_id);
return SUCCESS;
}
Status DavinciModel::InitStreamSwitchN(const OpDescPtr &op_desc) {
std::vector<uint32_t> active_stream_list;
if (!AttrUtils::GetListInt(op_desc, ATTR_NAME_ACTIVE_STREAM_LIST, active_stream_list)) {
GELOGE(INTERNAL_ERROR, "StreamSwitchNOp get attr ACTIVE_STREAM failed.");
return INTERNAL_ERROR;
}
for (size_t j = 0; j < active_stream_list.size(); ++j) {
active_stream_indication_.insert(active_stream_list[j]);
GELOGI("StreamSwitchNOp node:%s, active_stream_id=%u.", op_desc->GetName().c_str(), active_stream_list[j]);
}
uint32_t batch_num = 0;
if (!AttrUtils::GetInt(op_desc, ATTR_NAME_BATCH_NUM, batch_num)) {
GELOGE(FAILED, "Failed to get attr ATTR_NAME_BATCH_NUM, StreamSwitchN: %s.", op_desc->GetName().c_str());
return FAILED;
}
return SetDynamicBatchInfo(op_desc, batch_num);
}
Status DavinciModel::SetDynamicBatchInfo(const OpDescPtr &op_desc, uint32_t batch_num) {
batch_info_.clear();
combined_batch_info_.clear();
(void)AttrUtils::GetInt(op_desc, ATTR_DYNAMIC_TYPE, dynamic_type_);
(void)AttrUtils::GetListStr(op_desc, ATTR_USER_DESIGNEATE_SHAPE_ORDER, user_designate_shape_order_);
for (uint32_t i = 0; i < batch_num; ++i) {
std::vector<int64_t> batch_shape;
const std::string attr_name = ATTR_NAME_PRED_VALUE + "_" + std::to_string(i);
if (!AttrUtils::GetListInt(op_desc, attr_name, batch_shape)) {
GELOGE(FAILED, "Get attr ATTR_NAME_PRED_VALUE failed, Node: %s", op_desc->GetName().c_str());
batch_info_.clear();
return FAILED;
}
batch_info_.emplace_back(batch_shape);
batch_shape.clear();
const string attr_combined_batch = ATTR_NAME_COMBINED_BATCH + "_" + std::to_string(i);
if (AttrUtils::GetListInt(op_desc, attr_combined_batch, batch_shape)) {
combined_batch_info_.emplace_back(batch_shape);
}
}
return SUCCESS;
}
Status DavinciModel::InitCase(const OpDescPtr &op_desc) {
uint32_t batch_num = 0;
if (!AttrUtils::GetInt(op_desc, ATTR_NAME_BATCH_NUM, batch_num)) {
GELOGI("Not multi-batch Node: %s", op_desc->GetName().c_str());
return SUCCESS;
}
return SetDynamicBatchInfo(op_desc, batch_num);
}
bool DavinciModel::IsBroadCastOpData(const ge::NodePtr &var_node) {
for (auto out_anchor : var_node->GetAllOutDataAnchors()) {
GE_RT_FALSE_CHECK_NOTNULL(out_anchor);
for (auto in_anchor : out_anchor->GetPeerInDataAnchors()) {
GE_RT_FALSE_CHECK_NOTNULL(in_anchor);
ge::NodePtr dst_node = in_anchor->GetOwnerNode();
GE_RT_FALSE_CHECK_NOTNULL(dst_node);
if (dst_node->GetType() == HCOMBROADCAST || dst_node->GetType() == HVDCALLBACKBROADCAST) {
return true;
}
}
}
return false;
}
///
/// @ingroup ge
/// @brief Init model stream for NN model.
/// @param [in] stream user input model stream.
/// @return Status
///
Status DavinciModel::InitModelStream(rtStream_t stream) {
ExecuteMode curr_mode = is_async_mode_ ? ASYNCHRONIZATION : SYNCHRONIZATION;
GE_CHK_BOOL_RET_STATUS((curr_mode == last_execute_mode_) || (last_execute_mode_ == INITIALIZATION), INTERNAL_ERROR,
"NnExecute not support mix execute.");
last_execute_mode_ = curr_mode;
// asynchronize mode, use user input stream.
if (is_async_mode_) {
rt_model_stream_ = stream;
is_inner_model_stream_ = false;
return SUCCESS;
}
// synchronize mode, use forbidden stream.
if (stream != nullptr) {
if ((rt_model_stream_ != nullptr) && is_inner_model_stream_) {
GE_LOGW_IF(rtStreamDestroy(rt_model_stream_) != RT_ERROR_NONE, "Destroy rt_stream failed!");
}
rt_model_stream_ = stream;
is_inner_model_stream_ = false;
return SUCCESS;
}
if (rt_model_stream_ == nullptr) {
GE_CHK_RT_RET(rtStreamCreateWithFlags(&rt_model_stream_, priority_, RT_STREAM_FORBIDDEN_DEFAULT));
is_inner_model_stream_ = true;
}
return SUCCESS;
}
///
/// @ingroup ge
/// @brief ACL case, do not start new thread, return execute result.
/// @param [in] stream execute model stream.
/// @param [in] async_mode is asynchronize mode.
/// @param [in] input_data model input data.
/// @param [out] output_data model output data.
///
Status DavinciModel::NnExecute(rtStream_t stream, bool async_mode, const InputData &input_data,
OutputData &output_data) {
is_async_mode_ = async_mode;
GELOGI("Model Run begin, model id:%u, data index:%u, flag:%d.", model_id_, input_data.index, is_async_mode_);
GE_CHK_STATUS_RET(InitModelStream(stream), "Init model stream failed.");
is_dynamic_ = input_data.is_dynamic_batch;
if (!is_dynamic_) {
zero_copy_batch_label_addrs_.clear();
}
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_PRE_PROC_START));
Status ret = CopyModelData(input_data, output_data, is_dynamic_);
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(ret != SUCCESS, return ret, "Copy input data to model failed. model id: %u",
model_id_);
GELOGI("current_data.index=%u", input_data.index);
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_PRE_PROC_END));
if (!task_list_.empty()) {
GELOGD("rtModelExecute do");
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_INFER_START));
rtError_t rt_ret = rtModelExecute(rt_model_handle_, rt_model_stream_, 0);
GE_CHK_RT_EXEC(rt_ret, return RT_ERROR_TO_GE_STATUS(rt_ret));
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_INFER_END));
GELOGI("rtModelExecute end");
}
if (!is_async_mode_) {
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_AFTER_PROC_START));
ret = CopyOutputData(input_data.index, output_data, RT_MEMCPY_DEVICE_TO_DEVICE);
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(ret != SUCCESS, return ret, "Copy Output data to user failed.");
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), SetProfileTime(MODEL_AFTER_PROC_END));
}
// report model time data
GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), (void)SinkTimeProfile(input_data));
GELOGI("Model run end, model id:%u", model_id_);
return SUCCESS;
}
// Add active entry stream for special env.
Status DavinciModel::AddHeadStream() {
if (active_stream_list_.empty()) {
GELOGE(INTERNAL_ERROR, "Active stream is empty, stream list size: %zu, stream indication size: %zu.",
stream_list_.size(), active_stream_indication_.size());
return INTERNAL_ERROR;
}
if (active_stream_list_.size() == 1) {
GELOGI("Just one active stream, take as head stream.");
rt_head_stream_ = active_stream_list_[0];
is_pure_head_stream_ = false;
} else {
// Create stream which rt_model_handel running on, this is S0, TS stream.
GELOGI("Multiple active stream: %zu, create head stream.", active_stream_list_.size());
GE_CHK_RT_RET(rtStreamCreateWithFlags(&rt_head_stream_, priority_, RT_STREAM_PERSISTENT));
GE_CHK_RT_RET(rtModelBindStream(rt_model_handle_, rt_head_stream_, RT_INVALID_FLAG)); // Not active.
is_pure_head_stream_ = true;
for (auto s : active_stream_list_) {
std::shared_ptr<CpuTaskActiveEntry> active_entry = MakeShared<CpuTaskActiveEntry>(rt_head_stream_);
if (active_entry == nullptr) {
GELOGE(MEMALLOC_FAILED, "Make CpuTaskActiveEntry task failed.");
return MEMALLOC_FAILED;
}
Status status = active_entry->Init(s);
if (status != SUCCESS) {
return status;
}
cpu_task_list_.emplace_back(active_entry);
}
}
// Create entry stream active head stream. AICPU stream.
GE_CHK_RT_RET(rtStreamCreateWithFlags(&rt_entry_stream_, priority_, RT_STREAM_AICPU));
GE_CHK_RT_RET(rtModelBindStream(rt_model_handle_, rt_entry_stream_, RT_HEAD_STREAM));
return SUCCESS;
}
Status DavinciModel::InitEntryTask() {
if (deploy_type_ == AICPU_DEPLOY_CROSS_THREAD) {
GE_CHK_STATUS_RET(AddHeadStream(), "Add head stream failed.");
return CpuActiveStream();
} else {
return LoadWithQueue();
}
}
uint8_t *DavinciModel::MallocFeatureMapMem(size_t data_size) {
uint8_t *mem_base = nullptr;
const string purpose("feature map,used for op input and output.");
if (std::getenv(kEnvGeuseStaticMemory) != nullptr) {
data_size = static_cast<size_t>(VarManager::Instance(session_id_)->GetGraphMemoryMaxSize());
string memory_key = std::to_string(0) + "_f";
mem_base = MemManager::Instance(RT_MEMORY_HBM)->MallocMemory(purpose, memory_key, data_size, GetDeviceId());
} else {
mem_base = MemManager::Instance(RT_MEMORY_HBM)->MallocMemory(purpose, data_size, GetDeviceId());
}
if (mem_base != nullptr) {
GE_CHK_RT(rtMemset(mem_base, data_size, 0U, data_size));
}
return mem_base;
}
uint8_t *DavinciModel::MallocP2PMem(size_t p2p_data_size) {
uint8_t *p2p_mem_base = nullptr;
const string purpose("p2p memory, used for some op related to hcom");
if (std::getenv(kEnvGeuseStaticMemory) != nullptr) {
string p2p_memory_key = std::to_string(0) + "_p";
p2p_mem_base =
MemManager::Instance(RT_MEMORY_P2P_DDR)->MallocMemory(purpose, p2p_memory_key, p2p_data_size, GetDeviceId());
} else {
p2p_mem_base = MemManager::Instance(RT_MEMORY_P2P_DDR)->MallocMemory(purpose, p2p_data_size, GetDeviceId());
}
return p2p_mem_base;
}
uint8_t *DavinciModel::MallocWeightsMem(size_t weights_size) {
uint8_t *weights_mem_base = nullptr;
const string purpose("weights memory in inference network.");
if (std::getenv(kEnvGeuseStaticMemory) != nullptr) {
string weight_memory_key = std::to_string(0) + "_w";
weights_mem_base =
MemManager::Instance(RT_MEMORY_HBM)->MallocMemory(purpose, weight_memory_key, weights_size, GetDeviceId());
} else {
weights_mem_base = MemManager::Instance(RT_MEMORY_HBM)->MallocMemory(purpose, weights_size, GetDeviceId());
}
return weights_mem_base;
}
void DavinciModel::FreeFeatureMapMem() {
if (std::getenv(kEnvGeuseStaticMemory) != nullptr && is_inner_mem_base_) {
string weight_memory_key = std::to_string(0) + "_f";
if (MemManager::Instance(RT_MEMORY_HBM)->GetMemoryAddr(weight_memory_key) != nullptr) {
GE_CHK_STATUS(MemManager::Instance(RT_MEMORY_HBM)->FreeMemory(weight_memory_key, GetDeviceId()),
"failed to free weight memory");
}
mem_base_ = nullptr;
} else {
GE_IF_BOOL_EXEC(mem_base_ != nullptr && is_inner_mem_base_,
GE_CHK_STATUS(MemManager::Instance(RT_MEMORY_HBM)->FreeMemory(mem_base_, GetDeviceId()),
"failed to free feature_map memory");
mem_base_ = nullptr);
}
}
void DavinciModel::FreeP2PMem() {
if (std::getenv(kEnvGeuseStaticMemory) != nullptr) {
std::string p2p_memory_key = std::to_string(0) + "_p";
if (MemManager::Instance(RT_MEMORY_P2P_DDR)->GetMemoryAddr(p2p_memory_key) != nullptr) {
GE_CHK_STATUS(MemManager::Instance(RT_MEMORY_P2P_DDR)->FreeMemory(p2p_memory_key, GetDeviceId()),
"failed to free p2p memory");
}
p2p_mem_base_ = nullptr;
} else {
GE_IF_BOOL_EXEC(p2p_mem_base_ != nullptr && is_inner_mem_base_,
GE_CHK_STATUS(MemManager::Instance(RT_MEMORY_P2P_DDR)->FreeMemory(p2p_mem_base_, GetDeviceId()),
"failed to free p2p memory");
p2p_mem_base_ = nullptr);
}
}
void DavinciModel::FreeWeightsMem() {
if (std::getenv(kEnvGeuseStaticMemory) != nullptr) {
string memory_key = std::to_string(0) + "_w";
if (MemManager::Instance(RT_MEMORY_HBM)->GetMemoryAddr(memory_key) != nullptr) {
GE_CHK_STATUS(MemManager::Instance(RT_MEMORY_HBM)->FreeMemory(memory_key, GetDeviceId()),
"failed to free feature_map memory");
}
weights_mem_base_ = nullptr;
} else {
GE_IF_BOOL_EXEC(weights_mem_base_ != nullptr && weights_mem_base_ != mem_base_ && is_inner_weight_base_,
GE_CHK_STATUS(MemManager::Instance(RT_MEMORY_HBM)->FreeMemory(weights_mem_base_, GetDeviceId()),
"failed to free weight memory");
weights_mem_base_ = nullptr);
}
}
Status DavinciModel::TransAllVarData(ComputeGraphPtr &graph, uint32_t graph_id) {
GELOGI("TransAllVarData start: session_id:%lu, graph_id: %u.", session_id_, graph_id);
rtContext_t ctx = nullptr;
rtError_t rt_ret = rtCtxGetCurrent(&ctx);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Failed to get current context, error_code is: 0x%X.", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
std::vector<NodePtr> variable_node_list;
for (ge::NodePtr &node : graph->GetDirectNode()) {
if (node == nullptr) {
continue;
}
if (node->GetType() != VARIABLE) {
continue;
}
variable_node_list.emplace_back(node);
}
GE_CHK_STATUS_RET_NOLOG(
TransVarDataUtils::TransAllVarData(variable_node_list, session_id_, ctx, graph_id, kThreadNum));
GELOGI("TransAllVarData success.");
return SUCCESS;
}
void DavinciModel::SetDataDumperArgs(const ComputeGraphPtr &compute_graph) {
GELOGI("set data dumper args, name: %s, id: %u.", name_.c_str(), model_id_);
data_dumper_.SetModelName(name_);
data_dumper_.SetModelId(model_id_);
data_dumper_.SetOmName(om_name_);
data_dumper_.SetComputeGraph(compute_graph);
data_dumper_.SetRefInfo(saved_task_addrs_);
data_dumper_.SetL1FusionAddr(l1_fusion_addr_);
int32_t device_id = 0;
rtError_t rt_ret = rtGetDevice(&device_id);
if (rt_ret != RT_ERROR_NONE || device_id < 0) {
GELOGE(RT_FAILED, "Call rtGetDevice failed, ret = 0x%X, device_id = %d.", rt_ret, device_id);
return;
}
data_dumper_.SetDeviceId(device_id);
// set loop count addr
auto get_var_addr = [](const OpDescPtr &op, const RuntimeParam &runtime_param) -> void * {
if (op != nullptr) {
auto v_output_size = ModelUtils::GetOutputSize(op);
auto v_output_addr = ModelUtils::GetOutputDataAddrs(runtime_param, op);
if (v_output_size.empty() || v_output_addr.empty()) {
return nullptr;
}
return v_output_addr[0];
}
GELOGW("op is null.");
return nullptr;
};
data_dumper_.SetLoopAddr(get_var_addr(GetVariableOp(NODE_NAME_GLOBAL_STEP), runtime_param_),
get_var_addr(GetVariableOp(NODE_NAME_FLOWCTRL_LOOP_PER_ITER), runtime_param_),
get_var_addr(GetVariableOp(NODE_NAME_FLOWCTRL_LOOP_COND), runtime_param_));
GELOGI("SetDataDumperArgs end.");
}
uint32_t DavinciModel::GetFlowctrlIndex(uint32_t op_index) {
std::lock_guard<std::mutex> lock(flowctrl_op_index_internal_map_mutex_);
return (++flowctrl_op_index_internal_map_[op_index]) - 1;
}
void DavinciModel::PushHcclStream(rtStream_t value) {
std::lock_guard<std::mutex> lock(all_hccl_stream_list_mutex_);
all_hccl_stream_list_.push_back(value);
}
void DavinciModel::SaveHcclFollowStream(int64_t main_stream_id, rtStream_t stream) {
std::lock_guard<std::mutex> lock(capacity_of_stream_mutex_);
main_follow_stream_mapping_[main_stream_id].emplace_back(stream);
}
Status DavinciModel::GetComputeGraphInfo(vector<ComputeGraphDescInfo> &graph_desc_info) {
GELOGI("GetComputeGraphInfo start.");
auto &all_op_desc = data_dumper_.GetAllOpDescInfo();
for (auto &op_desc : all_op_desc) {
ComputeGraphDescInfo compute_graph_info;
if (!om_name_.empty()) {
compute_graph_info.model_name = om_name_;
} else {
compute_graph_info.model_name = name_;
}
compute_graph_info.op_name = op_desc.op_name;
compute_graph_info.op_type = op_desc.op_type;
compute_graph_info.input_format = op_desc.input_format;
compute_graph_info.input_shape = op_desc.input_shape;
compute_graph_info.input_data_type = op_desc.input_data_type;
compute_graph_info.output_format = op_desc.output_format;
compute_graph_info.output_shape = op_desc.output_shape;
compute_graph_info.output_data_type = op_desc.output_data_type;
graph_desc_info.emplace_back(compute_graph_info);
}
GELOGI("GetComputeGraphInfo end.");
return SUCCESS;
}
void DavinciModel::SetTotalFixedAddrsSize(string tensor_name, int64_t fix_addr_size) {
if (tensor_name_to_fixed_addr_size_.find(tensor_name) == tensor_name_to_fixed_addr_size_.end()) {
tensor_name_to_fixed_addr_size_[tensor_name] = total_fixed_addr_size_;
total_fixed_addr_size_ += fix_addr_size;
}
}
Status DavinciModel::GetOrigInputInfo(uint32_t index, OriginInputInfo &orig_input_info) {
GE_CHK_BOOL_RET_STATUS(index < data_op_list_.size(), PARAM_INVALID, "Index %u is invalid.", index);
OpDescPtr data_op = data_op_list_[index];
if (!data_op->HasAttr(ATTR_NAME_AIPP_INPUTS) || !data_op->HasAttr(ATTR_NAME_AIPP_OUTPUTS)) {
GELOGE(GE_AIPP_NOT_EXIST, "GetOrigInputInfo: there is not AIPP related with index %u.", index);
return GE_AIPP_NOT_EXIST;
}
vector<std::string> inputs;
if (AttrUtils::GetListStr(data_op, ATTR_NAME_AIPP_INPUTS, inputs) && !inputs.empty()) {
std::string input = inputs[kAippOriginInputIndex];
GELOGI("GetOrigInputInfo: origin input str: %s", input.c_str());
std::vector<std::string> infos = ge::StringUtils::Split(input, ':');
if (infos.size() != kAippInfoNum) {
GELOGW("origin input str is invalid.");
}
orig_input_info.format = TypeUtils::SerialStringToFormat(infos[kAippInfoFormat]);
orig_input_info.data_type = TypeUtils::SerialStringToDataType(infos[kAippInfoDataType]);
orig_input_info.dim_num = std::strtol(infos[kAippInfoDimNum].c_str(), nullptr, kDecimal);
}
return SUCCESS;
}
void DavinciModel::ParseAIPPInfo(std::string in_out_info, InputOutputDims &dims_info) {
GELOGI("ParseAIPPInfo: origin str: %s", in_out_info.c_str());
std::vector<std::string> infos = ge::StringUtils::Split(in_out_info, ':');
if (infos.size() != kAippInfoNum) {
GELOGW("origin input str is invalid.");
}
dims_info.name = infos[kAippInfoTensorName];
dims_info.size = std::strtol(infos[kAippInfoTensorSize].c_str(), nullptr, kDecimal);
dims_info.dim_num = std::strtol(infos[kAippInfoDimNum].c_str(), nullptr, kDecimal);
std::vector<std::string> dims = ge::StringUtils::Split(infos[kAippInfoShape], ',');
for (const auto &dim : dims) {
if (dim.empty()) {
continue;
}
dims_info.dims.emplace_back(std::strtol(dim.c_str(), nullptr, kDecimal));
}
}
Status DavinciModel::GetAllAippInputOutputDims(uint32_t index, std::vector<InputOutputDims> &input_dims,
std::vector<InputOutputDims> &output_dims) {
GE_CHK_BOOL_RET_STATUS(index < data_op_list_.size(), PARAM_INVALID, "Index %u is invalid.", index);
OpDescPtr data_op = data_op_list_[index];
if (!data_op->HasAttr(ATTR_NAME_AIPP_INPUTS) || !data_op->HasAttr(ATTR_NAME_AIPP_OUTPUTS)) {
GELOGE(GE_AIPP_NOT_EXIST, "GetAllAippInputOutputDims: there is not AIPP related with index %u.", index);
return GE_AIPP_NOT_EXIST;
}
vector<std::string> inputs;
if (AttrUtils::GetListStr(data_op, ATTR_NAME_AIPP_INPUTS, inputs) && !inputs.empty()) {
GELOGI("GetAllAippInputOutputDims: Data: %s has %u related aippInfo.", data_op->GetName().c_str(), inputs.size());
for (auto it : inputs) {
InputOutputDims input_info;
ParseAIPPInfo(it, input_info);
input_dims.emplace_back(input_info);
GELOGD("GetAllAippInputOutputDims Aipp origin input dims info: %s", it.c_str());
ConstGeTensorDescPtr data_input_desc = data_op->GetInputDescPtr(kDataIndex);
int64_t data_input_size;
(void)TensorUtils::GetSize(*(data_op->GetInputDescPtr(kDataIndex)), data_input_size);
GELOGD(
"GetAllAippInputOutputDims related Data[%d]: tensor_name is %s, dim_num is %u, tensor_size: %zu, format: "
"%s, data_type: %s, shape: %s .",
index, data_op->GetName().c_str(), data_input_desc->GetShape().GetDimNum(), data_input_size,
TypeUtils::FormatToSerialString(data_input_desc->GetFormat()).c_str(),
TypeUtils::DataTypeToSerialString(data_input_desc->GetDataType()).c_str(),
formats::JoinToString(data_input_desc->GetShape().GetDims()).c_str());
}
}
vector<std::string> outputs;
if (AttrUtils::GetListStr(data_op, ATTR_NAME_AIPP_OUTPUTS, outputs) && !outputs.empty()) {
for (auto it : outputs) {
InputOutputDims output_info;
ParseAIPPInfo(it, output_info);
output_dims.emplace_back(output_info);
GELOGD("GetAllAippInputOutputDims Aipp output dims info: %s", it.c_str());
}
}
return SUCCESS;
}
int64_t DavinciModel::GetFixedAddrsSize(string tensor_name) {
if (tensor_name_to_fixed_addr_size_.find(tensor_name) != tensor_name_to_fixed_addr_size_.end()) {
return tensor_name_to_fixed_addr_size_[tensor_name];
} else {
return total_fixed_addr_size_;
}
}
} // namespace ge
Reference in new issue View Git Blame Copy Permalink