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 '25aecfb52c'
${ noResults }
graphengine/ge/graph/build/memory/graph_mem_assigner.cc

1778 lines
83 KiB
Raw Blame History

/**
* Copyright 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/build/memory/graph_mem_assigner.h"
#include <cstring>
#include <set>
#include "common/math/math_util.h"
#include "common/util/error_manager/error_manager.h"
#include "framework/common/debug/ge_log.h"
#include "framework/common/debug/log.h"
#include "graph/build/memory/hybrid_mem_assigner.h"
#include "graph/build/memory/var_mem_assign_util.h"
#include "graph/build/memory/block_mem_assigner.h"
#include "graph/common/omg_util.h"
#include "graph/debug/ge_attr_define.h"
#include "graph/ge_attr_value.h"
#include "graph/manager/graph_var_manager.h"
#include "graph/utils/tensor_utils.h"
#include "graph/utils/type_utils.h"
#include "graph/build/memory/buffer_pool_mem_assigner.h"
namespace {
const int kAllInputAddrIsAtomic = -1;
const int kVirtualInputNodeMemoryReuse = 0;
const int kVirtualOutputNodeMemoryReuse = 1;
// One state per bit cannot be repeated
enum ContinuousType { kTypeInput = 1, kTypeInputNoPadding = 2, kTypeOutput = 4, kTypeOutputNoPadding = 8 };
int64_t GetSymbolOutputOffset(const std::map<std::string, std::string> &anchor_to_symbol,
const std::map<std::string, std::list<ge::NodeIndexIO>> &symbol_to_anchors,
const ge::NodePtr &node, const uint32_t i) {
ge::NodeIndexIO cur_node_index_io(node, i, ge::kOut);
auto iter1 = anchor_to_symbol.find(cur_node_index_io.ToString());
if (iter1 == anchor_to_symbol.end()) {
return ge::kInvalidOffset;
}
auto out_symbol = iter1->second;
auto iter2 = symbol_to_anchors.find(out_symbol);
if (iter2 == symbol_to_anchors.end()) {
return ge::kInvalidOffset;
}
for (const auto &node_index_io : iter2->second) {
if (node_index_io.value_ == out_symbol) {
vector<int64_t> output_list = node->GetOpDesc()->GetOutputOffset();
vector<int64_t> symbol_output_list = node_index_io.node_->GetOpDesc()->GetOutputOffset();
if (node_index_io.index_ >= symbol_output_list.size()) {
return ge::kInvalidOffset;
}
GELOGD("Node %s %uth output offset is %ld, Symbol %s output offset is %ld.", node->GetName().c_str(), i,
output_list[i], iter2->first.c_str(), symbol_output_list.at(node_index_io.index_));
return symbol_output_list.at(node_index_io.index_);
}
}
return ge::kInvalidOffset;
}
} // namespace
namespace ge {
Status VariableMemoryAssigner::Assign() {
Status result = ge::VarMemAssignUtil::AssignConstantOpMemory(compute_graph_);
if (result != ge::SUCCESS) {
return result;
}
result = ge::VarMemAssignUtil::AssignVarMemory(compute_graph_);
if (result != ge::SUCCESS) {
return result;
}
return ge::SUCCESS;
}
Status VariableMemoryAssigner::AssignVarAttr2Nodes() {
Status result = ge::VarMemAssignUtil::AssignVarAttr2Nodes(compute_graph_);
if (result != ge::SUCCESS) {
return result;
}
return ge::SUCCESS;
}
Status VariableMemoryAssigner::AssignMemory2HasRefAttrNode() {
Status result = ge::VarMemAssignUtil::AssignMemory2HasRefAttrNode(compute_graph_);
if (result != ge::SUCCESS) {
return result;
}
return ge::SUCCESS;
}
Status GraphMemoryAssigner::AssignMemory() {
ge::HybridMemAssignerPtr mem_assigner(new(std::nothrow) HybridMemAssigner(compute_graph_));
if (mem_assigner->Assign() != ge::SUCCESS) {
GELOGE(ge::FAILED, "[Assign][GraphMem]graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return ge::FAILED;
}
MemoryOffset memory_offset(RT_MEMORY_HBM, mem_assigner->GetMemOffset());
memory_offset_.emplace(RT_MEMORY_HBM, memory_offset);
if (mem_assigner->GetP2PMemOffset() >= 0) {
MemoryOffset p2p_memory_offset(RT_MEMORY_P2P_DDR, mem_assigner->GetP2PMemOffset());
memory_offset_.emplace(RT_MEMORY_P2P_DDR, p2p_memory_offset);
}
auto session_id = compute_graph_->GetSessionID();
int64_t var_size_before_assign = ge::VarManager::Instance(session_id)->GetVarMemSize(RT_MEMORY_HBM);
auto variable_assigner =
std::unique_ptr<ge::VariableMemoryAssigner>(new(std::nothrow) ge::VariableMemoryAssigner(compute_graph_));
if (variable_assigner == nullptr) {
GELOGE(ge::FAILED, "[New][Object:VariableMemoryAssigner]graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
REPORT_CALL_ERROR("E19999", "New Object:VariableMemoryAssigner failed, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return ge::FAILED;
}
if (variable_assigner->Assign() != ge::SUCCESS) {
return ge::FAILED;
}
int64_t var_size_assign = ge::VarManager::Instance(session_id)->GetVarMemSize(RT_MEMORY_HBM) - var_size_before_assign;
GELOGD("GraphMemoryAssigner::AssignMemory variable size = %ld", var_size_assign);
mem_assigner_ = std::move(mem_assigner);
return ge::SUCCESS;
}
ge::Status GraphMemoryAssigner::AssignVarAttr2Nodes() {
auto variable_assigner =
std::unique_ptr<ge::VariableMemoryAssigner>(new(std::nothrow) ge::VariableMemoryAssigner(compute_graph_));
if (variable_assigner == nullptr) {
GELOGE(ge::FAILED, "[New][Object:VariableMemoryAssigner]graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
REPORT_CALL_ERROR("E19999", "New Object:VariableMemoryAssigner failed, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return ge::FAILED;
}
if (variable_assigner->AssignVarAttr2Nodes() != ge::SUCCESS) {
return ge::FAILED;
}
return ge::SUCCESS;
}
ge::Status GraphMemoryAssigner::AssignMemory2HasRefAttrNode() {
auto variable_assigner =
std::unique_ptr<ge::VariableMemoryAssigner>(new(std::nothrow) ge::VariableMemoryAssigner(compute_graph_));
if (variable_assigner == nullptr) {
GELOGE(ge::FAILED, "[New][Object:VariableMemoryAssigner]graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
REPORT_CALL_ERROR("E19999", "New Object:VariableMemoryAssigner failed, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
}
if (variable_assigner->AssignMemory2HasRefAttrNode() != ge::SUCCESS) {
return ge::FAILED;
}
return ge::SUCCESS;
}
ge::Status CalculateTensorRealSizeAndOutSize(const ge::ConstGeTensorDescPtr &output_desc,
int64_t dim_index, int64_t &output_mem_size,
int64_t &batch_dim_num, int64_t &out_size) {
graphStatus graph_status = ge::TensorUtils::GetSize(*output_desc, out_size);
if (graph_status != GRAPH_SUCCESS) {
GELOGE(FAILED, "[Get][TensorSize]");
REPORT_CALL_ERROR("E19999", "Get tensor size failed");
return FAILED;
}
GeShape output_shape = output_desc->GetShape();
std::vector<int64_t> output_dims = output_shape.GetDims();
if (dim_index >= static_cast<int64_t>(output_dims.size())) {
REPORT_INNER_ERROR("E19999", "Inner param dim_index value:%ld invalid, bigger than dim size:%lu in shape:%s",
dim_index, output_dims.size(), output_shape.ToString().c_str());
GELOGE(FAILED, "[Check][Param:dim_index]value:%ld invalid, bigger than dim size:%lu in shape:%s",
dim_index, output_dims.size(), output_shape.ToString().c_str());
return FAILED;
}
for (int64_t index = 0; index < dim_index; index++) {
FMK_INT64_MULCHECK(batch_dim_num, output_dims[index]);
batch_dim_num *= output_dims[index];
output_dims[index] = 1;
}
output_shape = GeShape(output_dims);
Format out_format = output_desc->GetFormat();
DataType data_type = output_desc->GetDataType();
graph_status = ge::TensorUtils::CalcTensorMemSize(output_shape, out_format, data_type, output_mem_size);
if (graph_status != GRAPH_SUCCESS) {
GELOGE(graph_status, "[Calc][TensorSize]");
return FAILED;
}
if (output_mem_size < 0) {
REPORT_INNER_ERROR("E19999", "After calculating, tensor memory size:%ld invalid, less than 0. "
"shape:%s, format:%s, dtype:%s, maybe has dynamic shape",
output_mem_size,
output_shape.ToString().c_str(),
TypeUtils::FormatToSerialString(out_format).c_str(),
TypeUtils::DataTypeToSerialString(data_type).c_str());
GELOGE(FAILED, "[Check][TensorSize]value:%ld invalid after calc, less than 0. shape:%s, format:%s, dtype:%s, "
"maybe has dynamic shape",
output_mem_size,
output_shape.ToString().c_str(),
TypeUtils::FormatToSerialString(out_format).c_str(),
TypeUtils::DataTypeToSerialString(data_type).c_str());
return FAILED;
}
return SUCCESS;
}
Status GraphMemoryAssigner::ReAssignMemory(bool is_loop_graph, map<int64_t, size_t> &mem_type_to_offset) {
if (memory_offset_.empty()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ empty, not expected, graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData:memory_offset_]empty is not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return ge::FAILED;
}
GE_CHK_STATUS_RET(ReAssignContinuousMemory(is_loop_graph), "ReAssignContinuousMemory Failed!");
GE_CHK_STATUS_RET(ReAssignAtomicMemory(is_loop_graph), "ReAssignAtomicMemory Failed!");
GE_CHK_STATUS_RET(AssignBufferPoolMemory(), "AssignBufferPoolMemory Failed!");
size_t total_mem_offset = 0;
for (auto pair : memory_offset_) {
mem_type_to_offset[pair.first] = pair.second.mem_offset_;
total_mem_offset += pair.second.mem_offset_;
}
auto session_id = compute_graph_->GetSessionID();
if (total_mem_offset > VarManager::Instance(session_id)->GetGraphMemoryMaxSize()) {
GELOGE(ge::FAILED, "[Check][TotalMemOffset] %zu is greater than memory manager malloc max size %zu, "
"graph_id:%u, graph_name:%s, reduce your batchsize or scale your model may solve problem",
total_mem_offset, VarManager::Instance(session_id)->GetGraphMemoryMaxSize(),
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
for (auto iter : mem_type_to_offset) {
ErrorManager::GetInstance().ATCReportErrMessage("E19022", {"memType", "size", "item", "maxsize"},
{std::to_string(iter.first), std::to_string(iter.second), "featuremap",
std::to_string(VarManager::Instance(session_id)->GetGraphMemoryMaxSize())});
GEEVENT("[IMAS]AfterAssignMemory : %s memoffset[%zu], memtype[%ld]", compute_graph_->GetName().c_str(),
iter.second, iter.first);
}
return ge::FAILED;
}
return SUCCESS;
}
Status GraphMemoryAssigner::AssignZeroCopyMemory(map<int64_t, size_t> &mem_offset, size_t &zero_mem_copy_size) {
BlockMemAssignerPtr priority_assigner = std::move(mem_assigner_->GetPriorityAssinger());
if (priority_assigner == nullptr) {
REPORT_INNER_ERROR("E19999", "InnerData priority_assigner nullptr, not expected, graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData:priority_assigner]nullptr is invalid, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return ge::FAILED;
}
size_t mem_offset_tmp = mem_offset[RT_MEMORY_HBM];
// set offset for zero copy block
for (auto &memory_block : priority_assigner->GetMemoryBlocks()) {
if (memory_block == nullptr || memory_block->deleted_block_ || !memory_block->is_zero_copy_) {
continue;
}
memory_block->Resize();
memory_block->SetHeadOffset(mem_offset[RT_MEMORY_HBM]);
mem_offset[RT_MEMORY_HBM] += memory_block->Size();
memory_block->SetTailOffset(mem_offset[RT_MEMORY_HBM] - 1);
}
// set offset for zero copy nodes
priority_assigner->SetOpMemOffset(true);
zero_mem_copy_size = mem_offset[RT_MEMORY_HBM] - mem_offset_tmp;
auto iter = memory_offset_.find(RT_MEMORY_HBM);
if (iter == memory_offset_.end()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return FAILED;
}
iter->second.mem_offset_ = mem_offset[RT_MEMORY_HBM];
GELOGD("max_mem_offset:%zu, mem_offset:%zu, zero_mem_copy_size:%zu.", mem_offset[RT_MEMORY_HBM], mem_offset_tmp,
zero_mem_copy_size);
return SUCCESS;
}
uint32_t GetContinuousMemoryType(const OpDescPtr &op_desc) {
if (op_desc == nullptr) {
return 0;
};
bool is_continuous = false;
uint32_t continuous_type = 0;
// If GetBool fail, is_continuous is false.
(void)ge::AttrUtils::GetBool(op_desc, ATTR_NAME_CONTINUOUS_INPUT, is_continuous);
if (is_continuous) {
continuous_type |= kTypeInput;
} else {
(void)ge::AttrUtils::GetBool(op_desc, ATTR_NAME_NOPADDING_CONTINUOUS_INPUT, is_continuous);
if (is_continuous) {
bool attr_reuse = false;
(void)ge::AttrUtils::GetBool(op_desc, ATTR_NAME_OUTPUT_REUSE_INPUT, attr_reuse);
if (attr_reuse) {
continuous_type |= kTypeInputNoPadding;
}
}
}
is_continuous = false;
(void)ge::AttrUtils::GetBool(op_desc, ATTR_NAME_CONTINUOUS_OUTPUT, is_continuous);
if (is_continuous) {
continuous_type |= kTypeOutput;
} else {
(void)ge::AttrUtils::GetBool(op_desc, ATTR_NAME_NOPADDING_CONTINUOUS_OUTPUT, is_continuous);
if (is_continuous) {
bool attr_reuse = false;
(void)ge::AttrUtils::GetBool(op_desc, ATTR_NAME_OUTPUT_REUSE_INPUT, attr_reuse);
if (attr_reuse) {
continuous_type |= kTypeOutputNoPadding;
}
}
}
if (continuous_type != 0) {
GELOGI("[Get][MemType:Continuous]Current node %s, value is %d", op_desc->GetName().c_str(), continuous_type);
}
return continuous_type;
}
Status GetMemorySize(const OpDescPtr &op_desc, const ge::ConstGeTensorDescPtr &output_desc, uint32_t continuous_type,
int64_t &tensor_size, int64_t &nopadding_size) {
if ((op_desc == nullptr) || (output_desc == nullptr)) {
REPORT_INNER_ERROR("E19999", "InnerData param op_desc or output_desc is nullptr, not expected");
GELOGE(FAILED, "[Check][Param]op_desc or output_desc is nullptr");
}
tensor_size = 0;
nopadding_size = 0;
bool is_nopadding = ((continuous_type & kTypeInputNoPadding) != 0) || ((continuous_type & kTypeOutputNoPadding) != 0);
if (is_nopadding) {
int64_t attr_dim_index;
bool get_attr_dim_flag = ge::AttrUtils::GetInt(op_desc, ATTR_NAME_REUSE_INPUT_ON_DIM_INDEX, attr_dim_index);
if (!get_attr_dim_flag) {
REPORT_INNER_ERROR("E19999", "Get Attr:%s failed, op_name:%s",
ATTR_NAME_REUSE_INPUT_ON_DIM_INDEX.c_str(), op_desc->GetName().c_str());
GELOGE(FAILED, "[Get][Attr:%s]fail for op_name:%s",
ATTR_NAME_REUSE_INPUT_ON_DIM_INDEX.c_str(), op_desc->GetName().c_str());
return FAILED;
}
// Calculate tensor real size of each piece of data and out size of complete data
int64_t batch_dim_num = 1;
if (CalculateTensorRealSizeAndOutSize(output_desc, attr_dim_index, nopadding_size, batch_dim_num, tensor_size) !=
SUCCESS) {
REPORT_CALL_ERROR("E19999", "CalculateTensorRealSizeAndOutSize failed, attr_dim_index:%ld, op_name:%s",
attr_dim_index, op_desc->GetName().c_str());
GELOGE(FAILED, "[Calculate][NopaddingSize]failed for node %s, attr_dim_index:%ld",
op_desc->GetName().c_str(), attr_dim_index);
return FAILED;
}
} else {
if (ge::TensorUtils::GetSize(*output_desc, tensor_size) != ge::SUCCESS) {
REPORT_INNER_ERROR("E19999", "Get Tensor Size failed, op_name:%s", op_desc->GetName().c_str());
GELOGE(FAILED, "[Get][TensorSize]failed in padding case, op_name:%s", op_desc->GetName().c_str());
return FAILED;
}
}
if ((tensor_size < 0) || (nopadding_size < 0)) {
REPORT_INNER_ERROR("E19999", "GetMemorySize fail, "
"tensor_size:%ld or nopadding_size:%ld less than 0, invalid, op_name:%s",
tensor_size, nopadding_size, op_desc->GetName().c_str());
GELOGE(FAILED, "[Get][MemorySize]tensor_size:%ld or nopadding_size:%ld less than 0, invalid, op_name:%s",
tensor_size, nopadding_size, op_desc->GetName().c_str());
return FAILED;
}
return SUCCESS;
}
void AlignMemOffset(int64_t &mem_align_size) {
if (mem_align_size <= 0) {
return;
}
mem_align_size = (mem_align_size + MEM_ALIGN_SIZE - 1) / MEM_ALIGN_SIZE * MEM_ALIGN_SIZE;
}
bool IsContinuousInputConflict(const ge::NodePtr &node, const OpDescPtr &peer_op_desc) {
bool is_peer_output_continuous = false;
// If GetBool fail, is_peer_output_continuous is false.
(void) ge::AttrUtils::GetBool(peer_op_desc, ATTR_NAME_CONTINUOUS_OUTPUT, is_peer_output_continuous);
// Get peer node output size, if size == 1(peer node has only one output), continuous input of the node and
// continuous output of the previous node is the same, we can support it. If size != 1, there may be
// conflict between the two, we can not support it.
auto peer_output_size = peer_op_desc->GetOutputsSize();
GE_IF_BOOL_EXEC(is_peer_output_continuous && (peer_output_size != 1),
std::string error = "Current op" + FmtToStr(node->GetOpDesc()->GetName()) +
" requires continuous input, while the previous op" + FmtToStr(peer_op_desc->GetName()) +
" requires continuous output. There may be conflict between the two." +
"This node is not supported now.";
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return true;);
bool is_peer_reference = false;
// If GetBool fail, is_peer_reference is false.
(void) AttrUtils::GetBool(peer_op_desc, ATTR_NAME_REFERENCE, is_peer_reference);
GE_IF_BOOL_EXEC(is_peer_reference,
std::string warning = "[Check][Continuous]Current op" + FmtToStr(node->GetOpDesc()->GetName()) +
" requires continuous input, while the previous op" + FmtToStr(peer_op_desc->GetName()) +
" is ref. There may be conflict between the two.";
GELOGW("%s", warning.c_str());
return false;);
return false;
}
Status GraphMemoryAssigner::ReAssignContinuousMemory(bool is_loop_graph) {
Status ret;
// Stored nodes which need assign continuous input memory in `reverse topo order`
std::vector<NodePtr> nodes_stack;
std::map<NodePtr, uint32_t> node_2_continuous_type;
// Traverse nodes
for (auto &node : compute_graph_->GetAllNodes()) {
GE_CHECK_NOTNULL(node);
uint32_t continuous_type;
auto iter = node_2_continuous_type.find(node);
if (iter == node_2_continuous_type.end()) {
continuous_type = GetContinuousMemoryType(node->GetOpDesc());
node_2_continuous_type.emplace(node, continuous_type);
} else {
continuous_type = iter->second;
}
// Assign continuous input memory
bool continuous_input = ((continuous_type & kTypeInput) != 0) || ((continuous_type & kTypeInputNoPadding) != 0);
if (continuous_input) {
if (AssignContinuousInputMemoryWithAtomicProcessDirectly(node, node_2_continuous_type)) {
GE_CHK_STATUS_RET(AssignContinuousInputMemoryWithAtomicProcess(node, continuous_type),
"[Assign][Memory:Continuous:Input]fail for node:%s", node->GetName().c_str())
} else {
nodes_stack.push_back(node);
}
}
// Assign continuous output memory
int64_t memory_type = RT_MEMORY_HBM;
bool continuous_output = ((continuous_type & kTypeOutput) != 0) || ((continuous_type & kTypeOutputNoPadding) != 0);
if (continuous_output) {
GE_CHK_STATUS_RET(GetNodeMemoryType(node, memory_type, "output"),
"[Get][MemType]fail for node:%s", node->GetName().c_str());
ret = AssignContinuousOutputMemory(node, memory_type, continuous_type);
if (ret != ge::SUCCESS) {
GELOGE(ret, "[Assign][Memory:Continuous:Ouput]fail for node:%s", node->GetName().c_str());
return ret;
}
}
}
// Assign continuous input memory in `reverse topo order` which stored before
while (!nodes_stack.empty()){
auto node = nodes_stack.back();
nodes_stack.pop_back();
auto iter = node_2_continuous_type.find(node);
if (iter == node_2_continuous_type.end()) {
REPORT_INNER_ERROR("E19999", "Get ContinuousType from node_2_continuous_type map failed for node:%s",
node->GetName().c_str());
GELOGE(FAILED, "[Get][ContinuousType] find fail for node:%s", node->GetName().c_str());
return FAILED;
}
GE_CHK_STATUS_RET(AssignContinuousInputMemoryWithAtomicProcess(node, iter->second, true),
"[Assign][Memory:Continuous:Input]fail for node:%s.", node->GetName().c_str())
}
for (auto pair : memory_offset_) {
GELOGD("[Reassign][Memory:Continuous]At last, memory type = %ld, mem offset = %zu", pair.first,
pair.second.mem_offset_);
}
return ge::SUCCESS;
}
Status GraphMemoryAssigner::AssignContinuousInputMemory(const ge::NodePtr &node, int64_t &continuous_mem_start,
int64_t &continuous_mem_size, int64_t memory_type, uint32_t continuous_type, bool reverse_refresh) {
GELOGI("[Assign][Memory:Input:Continuous]start for Current node %s", node->GetName().c_str());
auto iter = memory_offset_.find(memory_type);
if (iter == memory_offset_.end()) {
REPORT_INNER_ERROR("E19999", "find memory offset fail for mem_type:%ld, "
"for node:%s, ", memory_type, node->GetName().c_str());
GELOGE(FAILED, "[Find][MemOffset]fail for mem_type:%ld, when AssignContinuousInputMemory for node:%s",
memory_type, node->GetName().c_str());
return FAILED;
}
// The head and tail of hcom continuous input should be added 512
iter->second.mem_offset_ += MEM_ALIGN_SIZE;
continuous_mem_start = iter->second.mem_offset_;
int64_t mem_offset = iter->second.mem_offset_;
int64_t extra_memory_size = 0;
bool is_continuous_input_allocated = false;
auto op_desc = node->GetOpDesc();
GE_CHECK_NOTNULL(op_desc);
vector<int64_t> output_list_this = op_desc->GetOutputOffset();
if (output_list_this.empty()) {
REPORT_INNER_ERROR("E19999", "No output offset in node :%s, not expected",
node->GetName().c_str());
GELOGE(FAILED, "[Get][OutputOffset] empty is invalid, node:%s", node->GetName().c_str());
return FAILED;
}
(void) ge::AttrUtils::GetBool(op_desc, ATTR_NAME_CONTINUOUS_INPUT_ALLOC, is_continuous_input_allocated);
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_op_desc = peer_out_data_anchor->GetOwnerNode()->GetOpDesc();
GE_IF_BOOL_EXEC(peer_op_desc == nullptr, continue);
GE_IF_BOOL_EXEC(IsContinuousInputConflict(node, peer_op_desc), return PARAM_INVALID;);
int64_t tensor_desc_size = 0;
int64_t nopadding_size = 0;
int64_t real_size = 0;
std::vector<int64_t> offsets_of_fusion = {};
bool lx_fusion = AttrUtils::GetListInt(peer_op_desc, ATTR_NAME_OUTPUT_OFFSET_FOR_BUFFER_FUSION, offsets_of_fusion);
lx_fusion = lx_fusion && !offsets_of_fusion.empty();
if (lx_fusion) {
if (peer_out_data_anchor->GetIdx() >= static_cast<int>(offsets_of_fusion.size())) {
std::string error = "fusion: peer node:" + FmtToStr(peer_op_desc->GetName()) +
" anchor_index:" + FmtToStr(peer_out_data_anchor->GetIdx()) +
" is out of range:" + FmtToStr(offsets_of_fusion.size());
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
nopadding_size = offsets_of_fusion[peer_out_data_anchor->GetIdx()];
tensor_desc_size = nopadding_size;
} else {
if (GetMemorySize(node->GetOpDesc(), peer_op_desc->GetOutputDescPtr(peer_out_data_anchor->GetIdx()),
continuous_type, tensor_desc_size, nopadding_size) != ge::SUCCESS) {
return FAILED;
}
}
bool is_nopadding = ((continuous_type & kTypeInputNoPadding) != 0) || lx_fusion;
vector<int64_t> output_list = peer_op_desc->GetOutputOffset();
if (peer_out_data_anchor->GetIdx() >= static_cast<int>(output_list.size())) {
std::string error = "peer node:" + FmtToStr(peer_op_desc->GetName()) +
" anchor_index:" + FmtToStr(peer_out_data_anchor->GetIdx()) +
" is out of range:" + FmtToStr(output_list.size());
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
// when continuous input has been allocated first input is beginning offset
bool is_allocated_first_input = is_continuous_input_allocated && (in_data_anchor->GetIdx() == 0);
if (is_allocated_first_input) {
std::map<int32_t, int32_t> out2ins;
GE_CHK_STATUS_RET(GetAllRef(node, out2ins), "[Get][AllRef]fail for node: %s", node->GetName().c_str());
// output is beginning offset, set offset for input; only support this case now
if ((out2ins.size() == 1) && (out2ins.begin()->second == 0) && (reverse_refresh)) {
auto peer_output_offset = output_list.at(peer_out_data_anchor->GetIdx());
output_list.at(peer_out_data_anchor->GetIdx()) = output_list_this.at(out2ins.begin()->first);
peer_op_desc->SetOutputOffset(output_list);
GELOGI("[Update][Offset]Node %s out %d ref in %d input node %s, use output offset %ld update %ld",
node->GetName().c_str(), out2ins.begin()->first, out2ins.begin()->second,
peer_op_desc->GetName().c_str(), output_list_this.at(out2ins.begin()->first), peer_output_offset);
} else {
GELOGD("Node %s out %d ref in %d input node %s with total ref numbers %zu.", node->GetName().c_str(),
out2ins.begin()->first, out2ins.begin()->second, peer_op_desc->GetName().c_str(), out2ins.size());
}
// first input is beginning offset
mem_offset = output_list.at(peer_out_data_anchor->GetIdx());
continuous_mem_start = output_list.at(peer_out_data_anchor->GetIdx());
} else {
// set offset for input
output_list.at(peer_out_data_anchor->GetIdx()) = mem_offset;
peer_op_desc->SetOutputOffset(output_list);
}
int64_t align_size = tensor_desc_size;
if (is_nopadding) {
mem_offset += nopadding_size;
extra_memory_size += (tensor_desc_size - nopadding_size);
real_size = nopadding_size;
} else {
ge::AlignMemOffset(align_size);
mem_offset += align_size;
// The head and tail of hcom continuous input should be added 512
extra_memory_size = MEM_ALIGN_SIZE;
real_size = tensor_desc_size;
}
GELOGI("[IMAS]Continuous input : Set %s name[%s] optype[%s] output[%d] offset to [%zu] stream_id[%ld] memtype[%ld] "
"size[%zu] realsize[%ld] nopadding size[%d]", node->GetOwnerComputeGraph()->GetName().c_str(),
peer_op_desc->GetName().c_str(), node->GetType().c_str(), peer_out_data_anchor->GetIdx(),
output_list.at(peer_out_data_anchor->GetIdx()), peer_op_desc->GetStreamId(), memory_type,
is_continuous_input_allocated ? 0UL : align_size, real_size, is_nopadding);
}
mem_offset += extra_memory_size;
ge::AlignMemOffset(mem_offset);
continuous_mem_size = mem_offset - continuous_mem_start;
if (is_continuous_input_allocated) {
// not allocate memory here, so no need add 512 in header
iter->second.mem_offset_ -= MEM_ALIGN_SIZE;
} else {
iter->second.mem_offset_ = mem_offset;
}
return SUCCESS;
}
Status GetFirstInputPeerOutOutputOffset(const ge::NodePtr &node, int64_t &mem_offset) {
auto in_data_anchor_list = node->GetAllInDataAnchors();
if (in_data_anchor_list.empty()) {
REPORT_INNER_ERROR("E19999", "InAnchor list empty in node:%s, not expect",
node->GetName().c_str());
GELOGE(FAILED, "[Get][InAnchor]empty is invalid, node:%s", node->GetName().c_str());
return FAILED;
}
auto peer_out_data_anchor = in_data_anchor_list.at(0)->GetPeerOutAnchor();
GE_IF_BOOL_EXEC(peer_out_data_anchor == nullptr,
REPORT_INNER_ERROR("E19999", "PeerAcnhor is null, not expect for node:%s",
node->GetName().c_str());
GELOGE(ge::FAILED, "[Check][PeerAnchor]null is invalid, node:%s", node->GetName().c_str());
return ge::FAILED);
auto peer_op_desc = peer_out_data_anchor->GetOwnerNode()->GetOpDesc();
GE_IF_BOOL_EXEC(peer_op_desc == nullptr,
REPORT_INNER_ERROR("E19999", "PeerOpDesc is null, not expect for node:%s",
node->GetName().c_str());
GELOGE(ge::FAILED, "[Check][PeerOpDesc]null is invalid, node:%s", node->GetName().c_str());
return ge::FAILED);
vector<int64_t> in_node_output_offsets = peer_op_desc->GetOutputOffset();
if (peer_out_data_anchor->GetIdx() >= static_cast<int>(in_node_output_offsets.size())) {
REPORT_INNER_ERROR("E19999", "PeerAnchorIndex:%d bigger than in_offset size:%lu, judge invalid for node:%s",
peer_out_data_anchor->GetIdx(), in_node_output_offsets.size(), node->GetName().c_str());
GELOGE(FAILED, "[Check][Index:PeerOutDataAnchor]PeerIndex:%d bigger than in_offset size:%lu, node:%s",
peer_out_data_anchor->GetIdx(), in_node_output_offsets.size(), node->GetName().c_str());
return FAILED;
}
mem_offset = in_node_output_offsets.at(peer_out_data_anchor->GetIdx());
return SUCCESS;
}
Status GraphMemoryAssigner::AssignContinuousOutputMemory(const ge::NodePtr &node, int64_t memory_type,
uint32_t continuous_type) {
GELOGI("Current node %s needs continuous output.", node->GetName().c_str());
auto out_op_desc = node->GetOpDesc();
GE_IF_BOOL_EXEC(out_op_desc == nullptr,
REPORT_INNER_ERROR("E19999", "OpDesc is null, not expect for node:%s",
node->GetName().c_str());
GELOGE(ge::FAILED, "[Check][OpDesc]null is invalid, node:%s", node->GetName().c_str()));
vector<int64_t> output_list = out_op_desc->GetOutputOffset();
if ((out_op_desc->GetOutputsSize() > output_list.size()) || (output_list.size() == 0)) {
REPORT_INNER_ERROR("E19999", "Output size:%zu more than output offset size:%zu, invalid in node:%s",
out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
GELOGE(ge::FAILED, "[Check][InnerData]Output size:%zu more than output offset size:%zu, invalid in node:%s",
out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
return ge::FAILED;
}
int64_t mem_offset = 0;
bool is_nopadding = ((continuous_type & kTypeOutputNoPadding) != 0);
if (is_nopadding) {
// out tensor memory must be reused input tensor memory
if (GetFirstInputPeerOutOutputOffset(node, mem_offset) != SUCCESS) {
return ge::FAILED;
}
} else {
// Get the reference type of the node, default is false
bool is_ref = false;
// If GetBool fail, is_ref is false.
(void) ge::AttrUtils::GetBool(node->GetOpDesc(), ATTR_NAME_REFERENCE, is_ref);
// If the output is ref type and refers to the ref of an input, the name of the output
// and the input are the same. Ge encounters ref type, finds matching relationship according
// to the names of input and output, and allocates the same memory address, eg: HCOMBroadcast
if (is_ref) {
GELOGI("Current node %s no needs assign continuous output because reference input by name.",
node->GetName().c_str());
return SUCCESS;
}
mem_offset = output_list[0];
}
for (auto &out_data_anchor : node->GetAllOutDataAnchors()) {
output_list[out_data_anchor->GetIdx()] = mem_offset;
int64_t tensor_desc_size = 0;
int64_t nopadding_size = 0;
if (GetMemorySize(out_op_desc, out_op_desc->GetOutputDescPtr(out_data_anchor->GetIdx()), continuous_type,
tensor_desc_size, nopadding_size) != ge::SUCCESS) {
return FAILED;
}
if (is_nopadding) {
mem_offset += nopadding_size;
} else {
mem_offset += tensor_desc_size;
ge::AlignMemOffset(mem_offset);
}
GELOGI("[IMAS]Continuous output : Set %s name[%s] optype[%s] output[%d] offset to [%zu] stream_id[%ld] memtype[%ld]"
" size[%zu] realsize[%ld] nopadding[%d].", node->GetOwnerComputeGraph()->GetName().c_str(),
out_op_desc->GetName().c_str(), node->GetType().c_str(), out_data_anchor->GetIdx(),
output_list[out_data_anchor->GetIdx()], out_op_desc->GetStreamId(), memory_type, 0UL,
is_nopadding ? nopadding_size : tensor_desc_size, is_nopadding);
}
out_op_desc->SetOutputOffset(output_list);
return ge::SUCCESS;
}
Status GraphMemoryAssigner::ReAssignAtomicMemory(bool is_loop_graph) {
// key:dynamic batch, batch name
map<string, map<NodePtr, vector<NodePtr>>> normal_atomic_and_clean_nodes_map;
map<string, vector<NodePtr>> connecting_output_atomic_nodes;
Status status = FilterAtomicNodesForMemoryAssign(normal_atomic_and_clean_nodes_map, connecting_output_atomic_nodes);
if (status != SUCCESS) {
GELOGE(status, "[Filter][AtomicNode]failed in graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return status;
}
auto mem_iter = memory_offset_.find(RT_MEMORY_HBM);
if (mem_iter == memory_offset_.end()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return FAILED;
}
int64_t batch_atomic_mem_start = static_cast<int64_t>(mem_iter->second.mem_offset_);
int64_t batch_max_mem_offset = batch_atomic_mem_start;
for (auto &iter_batch : normal_atomic_and_clean_nodes_map) {
mem_iter->second.mem_offset_ = batch_atomic_mem_start;
for (auto &iter : iter_batch.second) {
int64_t atomic_mem_start = static_cast<int64_t>(mem_iter->second.mem_offset_);
GELOGD("Begin to reAssign atomic memory, atomic address memory start = %ld", atomic_mem_start);
for (auto &atomic_node : iter.second) {
vector<int64_t> mem_offset_end;
status = AssignAtomicOutputAndWorkspaceMemory(atomic_node, mem_offset_end);
if (status != SUCCESS) {
GELOGE(status, "[Assign][Memory]output atomic mem and workspace mem, fail for node name is %s.",
atomic_node->GetName().c_str());
return status;
}
}
int64_t atomic_mem_size = static_cast<int64_t>(mem_iter->second.mem_offset_) - atomic_mem_start;
if (atomic_mem_size != 0) {
GE_CHK_STATUS_RET(SetAtomicCleanAttr(iter.first, {atomic_mem_start}, {atomic_mem_size}, RT_MEMORY_HBM),
"[Set][Attr]fail for atomic addr clean node %s.", iter.first->GetName().c_str());
}
}
batch_max_mem_offset = std::max(batch_max_mem_offset, static_cast<int64_t>(mem_iter->second.mem_offset_));
}
mem_iter->second.mem_offset_ = static_cast<size_t>(batch_max_mem_offset);
batch_atomic_mem_start = batch_max_mem_offset;
for (auto &iter_batch : connecting_output_atomic_nodes) {
mem_iter->second.mem_offset_ = batch_atomic_mem_start;
if (AssignConnectNetOutputAtomicMemory(iter_batch.second) != SUCCESS) {
GELOGE(FAILED, "[Assign][Memory]for nodes that connect to netoutput failed."
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return FAILED;
}
batch_max_mem_offset = std::max(batch_max_mem_offset, static_cast<int64_t>(mem_iter->second.mem_offset_));
}
mem_iter->second.mem_offset_ = static_cast<size_t>(batch_max_mem_offset);
return SUCCESS;
}
Status GraphMemoryAssigner::FilterAtomicNodesForMemoryAssign(
map<string, map<NodePtr, vector<NodePtr>>> &normal_atomic_nodes_map,
map<string, vector<NodePtr>> &connecting_output_atomic_nodes) {
GE_CHECK_NOTNULL(compute_graph_);
for (const auto &node : compute_graph_->GetAllNodes()) {
if (node->GetType() == ATOMICADDRCLEAN) {
map<string, vector<NodePtr>> tmp_normal_atomic_nodes;
const auto &out_control_anchor = node->GetOutControlAnchor();
GE_CHECK_NOTNULL(out_control_anchor);
for (const auto &peer_in_control_anchor : out_control_anchor->GetPeerInControlAnchors()) {
if (peer_in_control_anchor != nullptr) {
auto peer_in_node = peer_in_control_anchor->GetOwnerNode();
auto peer_in_node_desc = peer_in_node->GetOpDesc();
if (peer_in_node_desc != nullptr) {
bool is_atomic_node = false;
// If GetBool fail, is_atomic_node is false.
(void) ge::AttrUtils::GetBool(peer_in_node_desc, ATOMIC_ATTR_IS_ATOMIC_NODE, is_atomic_node);
if (is_atomic_node) {
bool is_reference = false;
// If GetBool fail, is_reference is false.
(void) ge::AttrUtils::GetBool(peer_in_node_desc, ATTR_NAME_REFERENCE, is_reference);
if (is_reference) {
REPORT_INNER_ERROR("E19999", "Op:%s cannot have both atomic and is_reference attribute, "
"not support now", peer_in_node_desc->GetName().c_str());
GELOGE(FAILED, "[Check][Attr]Op:%s cannot have both atomic and is_reference attribute, "
"not support now", peer_in_node_desc->GetName().c_str());
return ge::PARAM_INVALID;
}
std::string batch_label;
(void)ge::AttrUtils::GetStr(peer_in_node_desc, ATTR_NAME_BATCH_LABEL, batch_label);
vector<int> is_connecting_output;
// If GetBool fail, attr is_connecting_output is an empty vector.
(void) ge::AttrUtils::GetListInt(peer_in_node_desc, ATTR_NAME_NODE_CONNECT_OUTPUT, is_connecting_output);
if (is_connecting_output.empty()) {
tmp_normal_atomic_nodes[batch_label].emplace_back(peer_in_node);
continue;
}
connecting_output_atomic_nodes[batch_label].emplace_back(peer_in_node);
tmp_normal_atomic_nodes[batch_label].clear();
break;
}
}
}
}
for (auto &it_atomic_node : tmp_normal_atomic_nodes) {
if (!it_atomic_node.second.empty()) {
normal_atomic_nodes_map[it_atomic_node.first][node] = it_atomic_node.second;
}
}
}
}
return SUCCESS;
}
Status GraphMemoryAssigner::AssignAtomicOutputAndWorkspaceMemory(const ge::NodePtr &node,
vector<int64_t> &mem_offset_end) {
auto node_op_desc = node->GetOpDesc();
// Assign atomic node output memory
Status ret = AssignAtomicOutputMemory(node, mem_offset_end);
if (ret != SUCCESS) {
GELOGE(ret, "[Assign][Memory:Ouput:Atomic]Failed for node:%s.", node_op_desc->GetName().c_str());
return ret;
}
// Check and assign atomic node workspace memory
map<string, map<int64_t, int64_t>> atomic_workspace_info;
atomic_workspace_info = node_op_desc->TryGetExtAttr(EXT_ATTR_ATOMIC_WORKSPACE_INFO, atomic_workspace_info);
if (!atomic_workspace_info.empty()) {
bool is_fusion_node = false;
// If GetBool fail, is_fusion_node is false.
(void) ge::AttrUtils::GetBool(node_op_desc, ATOMIC_ATTR_IS_FUSION_NODE, is_fusion_node);
if (is_fusion_node) {
// Assign fusion atomic node workspace memory
ret = AssignFusionAtomicWorkspaceMemory(node_op_desc, atomic_workspace_info, mem_offset_end);
} else {
// Assign single ordinary atomic node workspace memory, not include fusion node
ret = AssignOrdinaryAtomicWorkspaceMemory(node_op_desc, atomic_workspace_info, mem_offset_end);
}
if (ret != SUCCESS) {
GELOGE(ret, "[Assign][Memory:Atomic:Workspace]fail for node:%s.", node_op_desc->GetName().c_str());
return ret;
}
} else {
GELOGW("Current atomic node %s does not have attr ATOMIC_WORKSPACE_INFO.", node->GetName().c_str());
}
return SUCCESS;
}
Status GraphMemoryAssigner::AssignConnectNetOutputAtomicMemory(vector<NodePtr> &connect_netoutput_nodes) {
auto iter = memory_offset_.find(RT_MEMORY_HBM);
if (iter == memory_offset_.end()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return FAILED;
}
for (auto &node : connect_netoutput_nodes) {
GE_CHECK_NOTNULL(node);
if (node->GetOpDesc() == nullptr) {
GELOGW("Current node %s op desc is nullptr, memory assignment is skipped.", node->GetName().c_str());
continue;
}
// Atomic memory start addr
int64_t original_atomic_mem_start = static_cast<int64_t>(iter->second.mem_offset_);
GELOGD("Start to assign memory of atomic node, node name: %s, node type: %s, mem_offset: %ld.",
node->GetName().c_str(), node->GetOpDesc()->GetType().c_str(), original_atomic_mem_start);
vector<int64_t> mem_offset_end;
if (AssignAtomicOutputAndWorkspaceMemory(node, mem_offset_end) != SUCCESS) {
GELOGE(FAILED, "[Assign][Memory]output atomic mem and workspace mem, fail for node name is %s.",
node->GetName().c_str());
return FAILED;
}
// All atomic nodes use atomic_addr_clean op independently, so we need to set the attr separately.
if (SetIndependentAtomicAttr(node, original_atomic_mem_start, mem_offset_end, RT_MEMORY_HBM) != SUCCESS) {
GELOGE(FAILED, "[Set][Attr:IndependentAtomic]fail for node:%s", node->GetName().c_str());
return FAILED;
}
}
return SUCCESS;
}
Status GraphMemoryAssigner::AssignReferenceMemory() {
for (auto &node : compute_graph_->GetDirectNode()) {
// Get the reference type of the node, default is false
bool is_ref = false;
// If GetBool fail, is_ref is false.
(void) ge::AttrUtils::GetBool(node->GetOpDesc(), ATTR_NAME_REFERENCE, is_ref);
if (!is_ref) {
continue;
}
GELOGI("Current node %s needs to support the reference relationship between output and input.",
node->GetName().c_str());
auto out_op_desc = node->GetOpDesc();
GE_IF_BOOL_EXEC(out_op_desc == nullptr, GELOGE(ge::FAILED, "out_op_desc is null."); return ge::FAILED);
vector<int64_t> output_list = out_op_desc->GetOutputOffset();
if (out_op_desc->GetOutputsSize() > output_list.size()) {
REPORT_INNER_ERROR("E19999", "Output size:%zu more than output offset size:%zu, judge invalid in node:%s",
out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
GELOGE(ge::FAILED, "[Check][InnerData]Output size:%zu more than output offset size:%zu, invalid in node:%s",
out_op_desc->GetOutputsSize(), output_list.size(), node->GetName().c_str());
return ge::FAILED;
}
map<string, int> input_name_index;
for (const auto &input_name : out_op_desc->GetAllInputNames()) {
int index = out_op_desc->GetInputIndexByName(input_name);
input_name_index.emplace(input_name, index);
}
for (auto &out_data_anchor : node->GetAllOutDataAnchors()) {
string out_data_anchor_name = out_op_desc->GetOutputNameByIndex(out_data_anchor->GetIdx());
auto iter = input_name_index.find(out_data_anchor_name);
if (iter != input_name_index.end()) {
int index = iter->second;
GELOGI("Reference memory: input anchor index = %d, input anchor name = %s, output anchor name = %s.", index,
iter->first.c_str(), out_data_anchor_name.c_str());
GE_CHECK_NOTNULL(node->GetInDataAnchor(index));
auto peer_out_anchor = node->GetInDataAnchor(index)->GetPeerOutAnchor();
GE_IF_BOOL_EXEC(peer_out_anchor == nullptr, continue);
int peer_out_anchor_index = peer_out_anchor->GetIdx();
auto peer_out_node = peer_out_anchor->GetOwnerNode();
auto peer_out_op_desc = peer_out_node->GetOpDesc();
GE_CHECK_NOTNULL(peer_out_op_desc);
output_list[out_data_anchor->GetIdx()] = peer_out_op_desc->GetOutputOffset()[peer_out_anchor_index];
GELOGI("Reference output : Set %s name[%s] output[%d] offset to [%ld] stream_id[%ld]",
node->GetOwnerComputeGraph()->GetName().c_str(), peer_out_op_desc->GetName().c_str(),
out_data_anchor->GetIdx(), output_list[out_data_anchor->GetIdx()], peer_out_op_desc->GetStreamId());
} else {
GELOGI("Reference output : origin %s name[%s] output[%d] offset is [%ld] stream_id[%ld]",
node->GetOwnerComputeGraph()->GetName().c_str(), out_op_desc->GetName().c_str(),
out_data_anchor->GetIdx(), output_list[out_data_anchor->GetIdx()], out_op_desc->GetStreamId());
}
}
out_op_desc->SetOutputOffset(output_list);
}
return ge::SUCCESS;
}
bool GraphMemoryAssigner::CheckInputIsSupportAtomic(const ge::NodePtr &node) {
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_op_desc = peer_out_data_anchor->GetOwnerNode()->GetOpDesc();
if (peer_op_desc == nullptr) {
continue;
}
if ((peer_op_desc->GetType() == CONSTANTOP) || (peer_op_desc->GetType() == AIPP_DATA_TYPE) ||
(peer_op_desc->GetType() == VARIABLE)) {
REPORT_INNER_ERROR("E19999", "node(type:%s, name:%s) link to atomic node(name:%s), "
"this situation not supported now",
peer_op_desc->GetType().c_str(), peer_op_desc->GetName().c_str(), node->GetName().c_str());
GELOGE(ge::FAILED, "[Check][Link]node(type:%s, name:%s) link to atomic node(name:%s), "
"this situation not supported now",
peer_op_desc->GetType().c_str(), peer_op_desc->GetName().c_str(), node->GetName().c_str());
return false;
}
}
return true;
}
Status GraphMemoryAssigner::AssignAtomicOutputMemory(const ge::NodePtr &node, vector<int64_t> &mem_offset_end) {
auto op_desc = node->GetOpDesc();
GE_IF_BOOL_EXEC(op_desc == nullptr, GELOGE(ge::FAILED, "op_desc is null."); return ge::FAILED);
mem_offset_end.clear();
GELOGD("Begin to assign atomic output memory, node = %s.", op_desc->GetName().c_str());
vector<int64_t> atomic_output_index;
// If GetListInt fail, atomic_output_index is empty.
(void) ge::AttrUtils::GetListInt(op_desc, ATOMIC_ATTR_OUTPUT_INDEX, atomic_output_index);
// Check atomic output
vector<int64_t> output_list = op_desc->GetOutputOffset();
if (atomic_output_index.size() > output_list.size()) {
std::string error =
"Op:" + FmtToStr(node->GetName()) + "'s size:" + FmtToStr(atomic_output_index.size()) +
" of atomic_output_index is more than the size:" + FmtToStr(output_list.size()) + " of output_list";
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return ge::FAILED;
}
auto output_list_size = static_cast<int64_t>(output_list.size());
auto iter = memory_offset_.find(RT_MEMORY_HBM);
if (iter == memory_offset_.end()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return FAILED;
}
for (auto &output_index : atomic_output_index) {
if (output_index >= output_list_size) {
std::string error =
"Op:" + FmtToStr(node->GetName()) + "'s atomic_output index:" + FmtToStr(output_index) +
" is more than the size:" + FmtToStr(output_list_size) + " of output_list.";
GE_ERRORLOG_AND_ERRORMSG(ge::PARAM_INVALID, error.c_str());
return ge::PARAM_INVALID;
}
// If the input of the cascade op needs to clear the atomic addr, there is no need to clear it separately here
bool is_assigned_mem = false;
if (GetMemoryAssignmentStatus(node, output_index, is_assigned_mem) != SUCCESS) {
GELOGE(ge::FAILED, "[Get][MemoryAssignmentStatus]fail for node %s, out_index:%ld",
node->GetName().c_str(), output_index);
return ge::FAILED;
}
// If you have already assigned an atomic address, skip it, and you don't need to reassign it.
if (is_assigned_mem) {
GELOGI(
"Node %s atomic output : we have assigned atomic memory as the input of next node in "
"ReAssignContinuousMemory function.",
op_desc->GetName().c_str());
continue;
}
auto output_desc = op_desc->GetAllOutputsDescPtr().at(output_index);
int64_t size = 0;
if (ge::TensorUtils::GetSize(*output_desc, size) != SUCCESS) {
GELOGI("Get size failed");
}
output_list[output_index] = iter->second.mem_offset_;
std::string batch_label;
(void)ge::AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
GELOGI("[IMAS]Atomic output : Set %s name[%s] optype[%s] output[%ld] offset to [%zu] stream_id[%ld] memtype[%u] "
"size[%ld] real_size[%ld] batch[%s].", compute_graph_->GetName().c_str(), op_desc->GetName().c_str(),
node->GetType().c_str(), output_index, iter->second.mem_offset_, op_desc->GetStreamId(), RT_MEMORY_HBM,
size, size, batch_label.c_str());
iter->second.mem_offset_ += size;
AlignMemOffset(MEM_ALIGN_SIZE, RT_MEMORY_HBM);
mem_offset_end.emplace_back(iter->second.mem_offset_);
}
op_desc->SetOutputOffset(output_list);
return ge::SUCCESS;
}
Status GraphMemoryAssigner::GetMemoryAssignmentStatus(const ge::NodePtr &node, int64_t output_index,
bool &is_mem_assigned) {
if (static_cast<size_t>(output_index) >= node->GetAllOutDataAnchors().size()) {
std::string error =
"Op:" + FmtToStr(node->GetName()) + "'s output index:" + FmtToStr(output_index) +
" is more than the size:" + FmtToStr(node->GetAllOutDataAnchors().size()) + " of node's AllOutDataAnchors.";
GE_ERRORLOG_AND_ERRORMSG(ge::PARAM_INVALID, error.c_str());
return ge::PARAM_INVALID;
}
auto out_data_anchor = node->GetAllOutDataAnchors().at(output_index);
GE_CHECK_NOTNULL(out_data_anchor);
auto input_anchors = out_data_anchor->GetPeerInDataAnchors();
for (auto &input_anchor : input_anchors) {
auto output_node = input_anchor->GetOwnerNode();
/// Get input atomic attr of peer output op, if atomic_input_index[0] = -1, indicates that the atomic address
/// has been assigned
vector<int64_t> atomic_input_index;
(void) ge::AttrUtils::GetListInt(output_node->GetOpDesc(), ATOMIC_ATTR_INPUT_INDEX, atomic_input_index);
if (!atomic_input_index.empty() && (atomic_input_index[0] == kAllInputAddrIsAtomic)) {
is_mem_assigned = true;
break;
}
}
return SUCCESS;
}
Status GraphMemoryAssigner::AssignOrdinaryAtomicWorkspaceMemory(const ge::OpDescPtr &op_desc,
map<string, map<int64_t, int64_t>> &workspace_info,
vector<int64_t> &mem_offset_end) {
GELOGI("Begin to reassign normal atomic memory, node = %s.", op_desc->GetName().c_str());
auto mem_type_iter = memory_offset_.find(RT_MEMORY_HBM);
if (mem_type_iter == memory_offset_.end()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return FAILED;
}
vector<int64_t> workspace_vector = op_desc->GetWorkspace();
for (auto iter = workspace_info.begin(); iter != workspace_info.end(); ++iter) {
if (op_desc->GetName() != iter->first) {
std::string error = "The node name" + FmtToStr(op_desc->GetName()) +
" and the node name" + FmtToStr(iter->first) + " in workspace info are inconsistent.";
GE_ERRORLOG_AND_ERRORMSG(ge::PARAM_INVALID, error.c_str());
return ge::PARAM_INVALID;
}
if (iter->second.empty()) {
continue;
}
for (auto &info_iter : iter->second) {
auto workspace_index = static_cast<uint64_t>(info_iter.first);
auto workspace_size = info_iter.second;
if (workspace_index >= workspace_vector.size()) {
std::string error = "The workspace index:" + FmtToStr(workspace_index) +
" is more than the size:" + FmtToStr(workspace_vector.size()) + " of workspace vector in op:" +
op_desc->GetName().c_str();
GE_ERRORLOG_AND_ERRORMSG(ge::PARAM_INVALID, error.c_str());
return ge::PARAM_INVALID;
}
workspace_vector[workspace_index] = mem_type_iter->second.mem_offset_;
std::string batch_label;
(void)ge::AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
GELOGI(
"[IMAS]Atomic ordinary workspace : Set %s name[%s] optype[%s] workspace[%lu] offset to [%zu] stream_id[%ld] "
"memtype[%u] size[%ld] real_size[%ld] batch[%s].",
compute_graph_->GetName().c_str(), op_desc->GetName().c_str(), op_desc->GetType().c_str(), workspace_index,
mem_type_iter->second.mem_offset_, op_desc->GetStreamId(), RT_MEMORY_HBM, workspace_size, workspace_size,
batch_label.c_str());
mem_type_iter->second.mem_offset_ += workspace_size;
mem_offset_end.emplace_back(mem_type_iter->second.mem_offset_);
}
}
op_desc->SetWorkspace(workspace_vector);
return SUCCESS;
}
Status GraphMemoryAssigner::AssignFusionAtomicWorkspaceMemory(const ge::OpDescPtr &op_desc,
map<string, map<int64_t, int64_t>> &workspace_info,
vector<int64_t> &mem_offset_end) {
GELOGI("Begin to reassign fusion atomic memory, node = %s.", op_desc->GetName().c_str());
auto mem_type_iter = memory_offset_.find(RT_MEMORY_HBM);
if (mem_type_iter == memory_offset_.end()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ does not have type[HBM], not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData]memory_offset_ does not have memory type[HBM]"
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
return FAILED;
}
map<string, map<int64_t, int64_t>> sub_node_workspace_offset;
for (auto &iter : workspace_info) {
if (iter.second.empty()) {
continue;
}
map<int64_t, int64_t> index_offset;
for (auto &info_iter : iter.second) {
auto workspace_index = static_cast<uint64_t>(info_iter.first);
auto workspace_size = info_iter.second;
size_t workspace_offset = mem_type_iter->second.mem_offset_;
std::string batch_label;
(void)ge::AttrUtils::GetStr(op_desc, ATTR_NAME_BATCH_LABEL, batch_label);
GELOGI(
"[IMAS]Atomic fusion workspace : Set %s name[%s] optype[%s] workspace[%lu] offset to [%zu] stream_id[%ld] "
"memtype[%u] ssize[%ld] real_size[%ld] batch[%s].", compute_graph_->GetName().c_str(),
op_desc->GetName().c_str(), op_desc->GetType().c_str(), workspace_index, mem_type_iter->second.mem_offset_,
op_desc->GetStreamId(), RT_MEMORY_HBM, workspace_size, workspace_size, batch_label.c_str());
mem_type_iter->second.mem_offset_ += workspace_size;
mem_offset_end.emplace_back(mem_type_iter->second.mem_offset_);
index_offset.insert(std::make_pair(workspace_index, workspace_offset));
}
sub_node_workspace_offset.insert(std::make_pair(iter.first, index_offset));
}
if (!(op_desc->SetExtAttr(EXT_ATTR_ATOMIC_WORKSPACE_OFFSET, sub_node_workspace_offset))) {
REPORT_INNER_ERROR("E19999", "Set Attr:%s fail for node:%s",
EXT_ATTR_ATOMIC_WORKSPACE_OFFSET.c_str(), op_desc->GetName().c_str());
GELOGE(FAILED, "[Set][Attr:%s]fail for node:%s.",
EXT_ATTR_ATOMIC_WORKSPACE_OFFSET.c_str(), op_desc->GetName().c_str());
return FAILED;
}
return SUCCESS;
}
Status GraphMemoryAssigner::CheckOffset() {
std::map<std::string, std::string> anchor_to_symbol;
std::map<std::string, std::list<NodeIndexIO>> symbol_to_anchors;
if (GraphUtils::GetRefMapping(compute_graph_, symbol_to_anchors, anchor_to_symbol) != GRAPH_SUCCESS) {
REPORT_CALL_ERROR("E19999", "Get ref-mapping for graph %s failed", compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Get][RefMapping]fail for graph %s", compute_graph_->GetName().c_str());
return FAILED;
}
for (const ge::NodePtr &node : compute_graph_->GetAllNodes()) {
GE_CHECK_NOTNULL(node->GetOpDesc());
vector<int64_t> input_list = node->GetOpDesc()->GetInputOffset();
for (auto input : input_list) {
if (input == ge::kInvalidOffset) {
std::string error = "Invalid input offset" + FmtToStr(ge::kInvalidOffset) +
+ " in node" + FmtToStr(node->GetName());
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
}
bool need_update_output = false;
vector<int64_t> output_list = node->GetOpDesc()->GetOutputOffset();
for (uint32_t i = 0; i < output_list.size(); ++i) {
if (output_list[i] == ge::kInvalidOffset) {
std::string error = "Invalid output offset" + FmtToStr(ge::kInvalidOffset) +
+ " in node" + FmtToStr(node->GetName());
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
if (node->GetType() == IDENTITY || node->GetType() == READVARIABLEOP) {
auto symbol_offset = GetSymbolOutputOffset(anchor_to_symbol, symbol_to_anchors, node, i);
if (symbol_offset != ge::kInvalidOffset && output_list[i] != symbol_offset) {
output_list[i] = symbol_offset;
need_update_output = true;
}
}
}
if (need_update_output) {
node->GetOpDesc()->SetOutputOffset(output_list);
}
vector<int64_t> workspace_list = node->GetOpDesc()->GetWorkspace();
for (auto workspace : workspace_list) {
if (workspace == ge::kInvalidOffset) {
std::string error = "Invalid workspace" + FmtToStr(ge::kInvalidOffset) +
+ " in node" + FmtToStr(node->GetName());
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
}
}
return SUCCESS;
}
ge::Status GraphMemoryAssigner::SetInputOffset() {
if (memory_offset_.empty()) {
REPORT_INNER_ERROR("E19999", "InnerData memory_offset_ empty, not expected, graph_id:%u, graph_name:%s",
compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
GELOGE(FAILED, "[Check][InnerData:memory_offset_]empty is not expected, "
"graph_id:%u, graph_name:%s", compute_graph_->GetGraphID(), compute_graph_->GetName().c_str());
}
for (auto pair : memory_offset_) {
GEEVENT("[IMAS]AfterAssignMemory : %s memoffset[%zu], memtype[%ld]", compute_graph_->GetName().c_str(),
pair.second.mem_offset_, pair.first);
}
for (const ge::NodePtr &node : compute_graph_->GetAllNodes()) {
if (UpdateOpInputOffset(node) != ge::SUCCESS) {
GELOGE(ge::FAILED, "[Update][Offset:Input]fail for op:%s", node->GetName().c_str());
return ge::FAILED;
}
}
return ge::SUCCESS;
}
NodePtr GraphMemoryAssigner::GetKnownInputNode(const NodePtr &node) const {
if (!node->GetOpDesc()->HasAttr(ATTR_NAME_PARENT_NODE_INDEX)) {
return node;
}
if (NodeUtils::IsDynamicShape(node)) {
return node;
}
return NodeUtils::GetParentInput(node);
}
ge::Status GraphMemoryAssigner::UpdateConstArgsOffset(const NodePtr &node, vector<int64_t> &input_list) const {
uint32_t parent_index = 0;
if (!AttrUtils::GetInt(node->GetOpDesc(), ATTR_NAME_PARENT_NODE_INDEX, parent_index)) {
return SUCCESS;
}
// Subgraph Data Node, check for constant input.
std::string op_type;
const auto &in_node = NodeUtils::GetParentInput(node);
if (NodeUtils::GetConstOpType(in_node, op_type)) {
input_list = in_node->GetOpDesc()->GetOutputOffset();
node->GetOpDesc()->SetOutputOffset(input_list); // Set Data output same as const output.
return SUCCESS; // Constant input.
}
// Memory allocated for dynamic shape subgraph Data.
if (NodeUtils::IsDynamicShape(node)) {
return SUCCESS;
}
const auto &owner = node->GetOwnerComputeGraph();
const auto &parent_desc = owner->GetParentNode()->GetOpDesc();
const auto parent_inputs = parent_desc->GetInputOffset();
if (parent_inputs.size() <= parent_index) {
std::string error = "Get Parent input offset failed, node is " + FmtToStr(node->GetName()) +
+ ", input_size is " + FmtToStr(parent_inputs.size()) + ", parent index is " +
FmtToStr(parent_index);
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
input_list = {parent_inputs[parent_index]};
node->GetOpDesc()->SetOutputOffset(input_list); // Set Data output same as parent input.
return SUCCESS;
}
ge::Status GraphMemoryAssigner::UpdateOpInputOffset(const NodePtr &node, vector<int64_t> &input_list) const {
vector<int64_t> origin_input_list;
vector<int64_t> memory_type;
auto tmp_op_desc = node->GetOpDesc();
origin_input_list = tmp_op_desc->GetInputOffset();
int64_t valid_input_index = 0;
bool has_mem_type_attr = ge::AttrUtils::GetListInt(tmp_op_desc, ATTR_NAME_INPUT_MEM_TYPE_LIST, memory_type);
for (const auto &anchor : node->GetAllInDataAnchors()) {
vector<int64_t> output_list;
auto peer_out_anchor = anchor->GetPeerOutAnchor();
if (peer_out_anchor == nullptr) {
continue;
}
// If the current node not broadcast, the OutputOffset of the previous node is used to update the input_list
auto last_peer_out_node = peer_out_anchor->GetOwnerNode();
auto last_peer_out_op_desc = last_peer_out_node->GetOpDesc();
GE_CHECK_NOTNULL(last_peer_out_op_desc);
output_list = last_peer_out_op_desc->GetOutputOffset();
auto out_index = static_cast<unsigned long>(peer_out_anchor->GetIdx());
if (output_list.size() > static_cast<size_t>(out_index)) {
int64_t input_offset = output_list.at(out_index);
if (has_mem_type_attr && !origin_input_list.empty()) {
auto input_size = tmp_op_desc->GetInputsSize();
auto ori_input_offset_list_size = origin_input_list.size();
auto mem_type_size = memory_type.size();
if ((input_size != mem_type_size) || (input_size != ori_input_offset_list_size)) {
std::string error = "fusion: node" + FmtToStr(tmp_op_desc->GetName()) +
+ " input_size" + FmtToStr(input_size) + " diff from memory_type_size" +
FmtToStr(mem_type_size) + " from ori_input_offset_list_size" +
FmtToStr(ori_input_offset_list_size);
GE_ERRORLOG_AND_ERRORMSG(ge::FAILED, error.c_str());
return ge::FAILED;
}
// not hbm keep orignal inputoffest
// hbm inputoffset = original inputoffset + outputoffset
input_offset = (memory_type[valid_input_index] == RT_MEMORY_L1 ? origin_input_list[valid_input_index]
: origin_input_list[valid_input_index] + output_list.at(out_index));
}
const auto &in_node = GetKnownInputNode(peer_out_anchor->GetOwnerNode());
if (in_node->GetType() == CONSTANT) {
GeTensorDesc tensor_desc = tmp_op_desc->GetInputDesc(static_cast<uint32_t>(anchor->GetIdx()));
GE_CHK_STATUS(TensorUtils::GetDataOffset(tensor_desc, input_offset));
}
GELOGD("%s node[%s] input[%ld] is set from node[%s] out index[%lu] offset[%ld]",
has_mem_type_attr ? "Fusion" : "",
tmp_op_desc->GetName().c_str(),
valid_input_index,
peer_out_anchor->GetOwnerNode()->GetOpDesc()->GetName().c_str(),
out_index,
input_offset);
input_list.emplace_back(input_offset);
valid_input_index++;
}
}
return ge::SUCCESS;
}
ge::Status GraphMemoryAssigner::UpdateOpInputOffset(const NodePtr &node) const {
GE_CHECK_NOTNULL(node->GetOpDesc());
vector<int64_t> input_list;
if (node->GetType() == HCOMBROADCAST || node->GetType() == HVDCALLBACKBROADCAST) {
for (const auto &anchor : node->GetAllInDataAnchors()) {
vector<int64_t> output_list;
auto peer_out_anchor = anchor->GetPeerOutAnchor();
if (peer_out_anchor == nullptr) {
continue;
}
auto last_peer_out_node = peer_out_anchor->GetOwnerNode();
// If the current node is broadcast and the preceding node is variable, because InputOffset has been set
// in function:AssignVarAttr2Nodes, then the InputOffset of the broadcast node is taken to update the input_list.
// Otherwise, the OutputOffset of the previous node is used to update the input_list.
if (last_peer_out_node->GetType() != VARIABLE) {
auto last_peer_out_op_desc = last_peer_out_node->GetOpDesc();
GE_CHECK_NOTNULL(last_peer_out_op_desc);
output_list = last_peer_out_op_desc->GetOutputOffset();
if (output_list.size() > static_cast<size_t>(peer_out_anchor->GetIdx())) {
input_list.emplace_back(output_list.at(peer_out_anchor->GetIdx()));
}
} else {
vector<int64_t> cur_node_input_list;
auto cur_node_op_desc = node->GetOpDesc();
GE_CHECK_NOTNULL(cur_node_op_desc);
cur_node_input_list = cur_node_op_desc->GetInputOffset();
if (cur_node_input_list.size() > static_cast<size_t>(anchor->GetIdx())) {
input_list.emplace_back(cur_node_input_list.at(anchor->GetIdx()));
}
}
}
} else if (node->GetType() == DATA_TYPE) {
if (UpdateConstArgsOffset(node, input_list) != SUCCESS) {
GELOGE(FAILED, "[Update][Offset:Input:Const]fail for node:%s ", node->GetName().c_str());
return FAILED;
}
} else {
if (UpdateOpInputOffset(node, input_list) != SUCCESS) {
GELOGE(FAILED, "[Update][Offset:Input]fail for node:%s", node->GetName().c_str());
return FAILED;
}
}
node->GetOpDesc()->SetInputOffset(input_list);
return SUCCESS;
}
Status GraphMemoryAssigner::SetIndependentAtomicAttr(const ge::NodePtr &node, int64_t atomic_mem_start,
const vector<int64_t> &mem_offset_end, int64_t memory_type) {
GELOGD("Start to set independent atomic attr, atomic_addr_clean memory offset start is %ld", atomic_mem_start);
// Parsing offset and size vectors
vector<int64_t> memory_offset_start;
vector<int64_t> memory_offset_size;
memory_offset_start.emplace_back(atomic_mem_start);
for (size_t i = 0; i < mem_offset_end.size(); ++i) {
memory_offset_start.emplace_back(mem_offset_end[i]);
// Number 1 means element index
auto size = memory_offset_start[i + 1] - memory_offset_start[i];
memory_offset_size.emplace_back(size);
}
memory_offset_start.pop_back();
const auto &in_control_anchor = node->GetInControlAnchor();
if (!memory_offset_size.empty() && in_control_anchor != nullptr) {
for (auto &peer_out_control_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
if (peer_out_control_anchor == nullptr) {
continue;
}
auto peer_out_node = peer_out_control_anchor->GetOwnerNode();
auto peer_out_node_desc = peer_out_node->GetOpDesc();
if (peer_out_node_desc == nullptr) {
continue;
}
GELOGD("Current node memory_offset vector size is %zu, node name %s, node type is %s.", memory_offset_size.size(),
peer_out_node_desc->GetName().c_str(), peer_out_node_desc->GetType().c_str());
if (peer_out_node_desc->GetType() == ATOMICADDRCLEAN) {
if (SetAtomicCleanAttr(peer_out_node, memory_offset_start, memory_offset_size, memory_type) != SUCCESS) {
GELOGE(FAILED, "[Set][AtomicCleanAttr]fail for node:%s", peer_out_node->GetName().c_str());
return FAILED;
}
}
}
}
return SUCCESS;
}
ge::Status GraphMemoryAssigner::SetAtomicCleanAttr(const NodePtr &node, const vector<int64_t> &atomic_mem_start,
const vector<int64_t> &atomic_mem_size, int64_t memory_type) {
auto node_op_desc = node->GetOpDesc();
if (node_op_desc != nullptr) {
GELOGD("Node %s, set atomic clean attr start.", node->GetName().c_str());
vector<int64_t> workspace_vector = node_op_desc->GetWorkspace();
vector<int64_t> workspace_byte_vector = node_op_desc->GetWorkspaceBytes();
workspace_vector.insert(workspace_vector.end(), atomic_mem_start.begin(), atomic_mem_start.end());
workspace_byte_vector.insert(workspace_byte_vector.end(), atomic_mem_size.begin(), atomic_mem_size.end());
node_op_desc->SetWorkspace(workspace_vector);
node_op_desc->SetWorkspaceBytes(workspace_byte_vector);
std::vector<int64_t> mem_start_vector;
// If GetListInt fail, mem_start_vector is empty.
(void) ge::AttrUtils::GetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_START, mem_start_vector);
mem_start_vector.insert(mem_start_vector.end(), atomic_mem_start.begin(), atomic_mem_start.end());
GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_START, mem_start_vector),
REPORT_INNER_ERROR("E19999", "Set Attr:%s failed, op_name:%s",
ATTR_NAME_AUTOMIC_ADD_START.c_str(), node_op_desc->GetName().c_str());
GELOGE(FAILED, "[Set][Attr:%s]fail for op_name:%s",
ATTR_NAME_AUTOMIC_ADD_START.c_str(), node_op_desc->GetName().c_str());
return FAILED);
std::vector<int64_t> mem_size_vector;
// If GetListInt fail, mem_size_vector is empty.
(void) ge::AttrUtils::GetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_MEM_SIZE, mem_size_vector);
mem_size_vector.insert(mem_size_vector.end(), atomic_mem_size.begin(), atomic_mem_size.end());
GE_CHK_BOOL_EXEC(ge::AttrUtils::SetListInt(node_op_desc, ATTR_NAME_AUTOMIC_ADD_MEM_SIZE, mem_size_vector),
REPORT_INNER_ERROR("E19999", "Set Attr:%s failed, op_name:%s",
ATTR_NAME_AUTOMIC_ADD_MEM_SIZE.c_str(), node_op_desc->GetName().c_str());
GELOGE(FAILED, "[Set][Attr:%s]fail for op_name:%s",
ATTR_NAME_AUTOMIC_ADD_MEM_SIZE.c_str(), node_op_desc->GetName().c_str());
return FAILED);
std::stringstream ss;
for (auto iter : atomic_mem_start) {
ss << iter << " ";
}
string atomic_mem_start_str = ss.str();
ss.clear();
ss.str("");
for (auto iter : atomic_mem_size) {
ss << iter << " ";
}
string atomic_mem_size_str = ss.str();
GELOGI("[IMAS]SetAtomicCleanAttr : Set %s atomic_node name[%s] optype[%s] output[0] offset to [%s] streamid[%ld]"
" memtype[%ld] size[%s]",node->GetOwnerComputeGraph()->GetName().c_str(), node_op_desc->GetName().c_str(),
node->GetType().c_str(), atomic_mem_start_str.c_str(), node->GetOpDesc()->GetStreamId(), memory_type,
atomic_mem_size_str.c_str());
}
return SUCCESS;
}
void GraphMemoryAssigner::AlignMemOffset(const int64_t &mem_align_size, int64_t memory_type) {
if (mem_align_size <= 0) {
return;
}
auto iter = memory_offset_.find(memory_type);
if (iter == memory_offset_.end()) {
GELOGW("Memory offset don't have memory type[%ld].", memory_type);
return;
}
iter->second.mem_offset_ =
(iter->second.mem_offset_ + mem_align_size - 1) / mem_align_size * mem_align_size;
}
ge::Status GraphMemoryAssigner::GetNodeListMemoryType(const vector<NodePtr> &nodes, int32_t mem_reuse_model,
int64_t &memory_type) {
memory_type = RT_MEMORY_HBM;
// In the dynamic batch scenario, the memory attributes of nodes are the same.
for (auto &n : nodes) {
if (mem_reuse_model == kVirtualInputNodeMemoryReuse) {
GE_CHK_STATUS_RET(GetNodeMemoryType(n, memory_type, "input"),
"[Get][MemType:input]fail for node:%s", n->GetName().c_str())
break;
}
if (mem_reuse_model == kVirtualOutputNodeMemoryReuse) {
GE_CHK_STATUS_RET(GetNodeMemoryType(n, memory_type, "output"),
"[Get][MemType:output]fail for node:%s", n->GetName().c_str())
break;
}
}
return SUCCESS;
}
ge::Status GraphMemoryAssigner::GetNodeMemoryType(const NodePtr &node, int64_t &memory_type, string input_or_output) {
memory_type = RT_MEMORY_HBM;
vector<int64_t> mem_type_list;
if (input_or_output == "input") {
(void) ge::AttrUtils::GetListInt(node->GetOpDesc(), ATTR_NAME_INPUT_MEM_TYPE_LIST, mem_type_list);
}
if (input_or_output == "output") {
(void) ge::AttrUtils::GetListInt(node->GetOpDesc(), ATTR_NAME_OUTPUT_MEM_TYPE_LIST, mem_type_list);
}
if (mem_type_list.empty()) {
if (memory_offset_.find(memory_type) == memory_offset_.end()) {
std::string error = "Memory offset map does not have memory type" + FmtToStr(memory_type) +
+ ", opname is " + FmtToStr(node->GetName()) + ", optype is " + FmtToStr(node->GetType());
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
return SUCCESS;
}
if (mem_type_list.size() != node->GetAllInDataAnchorsSize()) {
std::string error = "The size" + FmtToStr(mem_type_list.size()) +
" of mem type list is not equal to the size of in data anchor" +
FmtToStr(node->GetAllInDataAnchorsSize()) + ", opname is " +
FmtToStr(node->GetName()) + ", optype is " + FmtToStr(node->GetType());
GE_ERRORLOG_AND_ERRORMSG(FAILED, error.c_str());
return FAILED;
}
if (!CheckContinuousMemType(mem_type_list)) {
GELOGE(FAILED, "[Check][MemType:Continuous]fail for node:%s", node->GetName().c_str());
return FAILED;
}
// It is continuous memory and memory type is the same, so use the first memory.
memory_type = mem_type_list[0];
return SUCCESS;
}
bool GraphMemoryAssigner::CheckContinuousMemType(vector<int64_t> mem_type_list) {
if (mem_type_list.size() == 0) {
return true;
}
int64_t mem_type_tmp = mem_type_list[0];
for (auto mem_type : mem_type_list) {
if (mem_type != mem_type_tmp) {
std::string error = "The memory is continuous, but the type of the input memory is inconsistent. They are " +
FmtToStr(mem_type_tmp) + " and " + FmtToStr(mem_type);
ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {error});
GELOGW("The memory is continuous, but the type of the input memory is inconsistent. They are [%ld] and [%ld].",
mem_type_tmp, mem_type);
return false;
}
}
if (memory_offset_.find(mem_type_tmp) == memory_offset_.end()) {
std::string error = "Memory offset map does not have memory type" + FmtToStr(mem_type_tmp);
ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {error});
GELOGW("Memory offset map does not have memory type[%ld].", mem_type_tmp);
return false;
}
return true;
}
void GraphMemoryAssigner::PrintMemoryOffset() {
for (auto pair : memory_offset_) {
// Assign memory of max batch nodes that have the same batch label.
GELOGD("Reassign memory for max batch virtual nodes, memory type = %ld, memory offset = %zu.",
pair.first, pair.second.mem_offset_);
}
}
ge::Status GraphMemoryAssigner::GetAllRef(const NodePtr &node, map<int32_t, int32_t> &out2ins) {
for (const auto &out_data_anchor : node->GetAllOutDataAnchors()) {
int32_t reuse_in_index = -1;
bool reuse_input_flag = GraphUtils::IsRefFromInput(out_data_anchor, reuse_in_index);
if (reuse_input_flag) {
if (node->GetInDataAnchor(reuse_in_index) != nullptr) {
out2ins.emplace(out_data_anchor->GetIdx(), reuse_in_index);
} else {
REPORT_INNER_ERROR("E19999", "Invalid reuse_input value %d on output %d of node %s, "
"please check attr reuse_input",
reuse_in_index, out_data_anchor->GetIdx(), node->GetName().c_str());
GELOGE(FAILED, "[Check][Attr]Invalid reuse_input value %d on output %d of node %s, "
"please check attr reuse_input",
reuse_in_index, out_data_anchor->GetIdx(), node->GetName().c_str());
return FAILED;
}
}
}
return ge::SUCCESS;
}
bool GraphMemoryAssigner::AssignContinuousInputMemoryWithAtomicProcessDirectly(
const NodePtr &input_continuous_node, map<NodePtr, uint32_t> &node_2_continuous_type) {
for (const auto &in_node : input_continuous_node->GetInDataNodes()) {
if (in_node->GetType() == VARIABLE) {
GELOGI("node %s 's precursor node %s is variable, do not store.", input_continuous_node->GetName().c_str(),
in_node->GetName().c_str());
return true;
}
auto iter = node_2_continuous_type.find(in_node);
// In node's topo order in the front, so function can not be exception
auto continuous_type = iter->second;
bool continuous_input = ((continuous_type & kTypeInput) != 0) || ((continuous_type & kTypeInputNoPadding) != 0);
if (continuous_input) {
GELOGI("[Store][Node] of %s cause it's precursor node %s need assign continuous input memory",
input_continuous_node->GetName().c_str(), in_node->GetName().c_str());
return false;
}
}
for (const auto &out_node : input_continuous_node->GetOutDataNodes()) {
auto continuous_type = GetContinuousMemoryType(out_node->GetOpDesc());
node_2_continuous_type.emplace(out_node, continuous_type);
bool continuous_input = ((continuous_type & kTypeInput) != 0) || ((continuous_type & kTypeInputNoPadding) != 0);
if (continuous_input) {
GELOGI("[Store][Node] of %s cause it's succeed node %s need assign continuous input memory",
input_continuous_node->GetName().c_str(), out_node->GetName().c_str());
return false;
}
}
return true;
}
ge::Status GraphMemoryAssigner::AssignContinuousInputMemoryWithAtomicProcess(const NodePtr &input_continuous_node,
uint32_t continuous_type,
bool reverse_refresh) {
int64_t mem_clean_start = 0;
int64_t mem_clean_size = 0;
int64_t memory_type = RT_MEMORY_HBM;
GE_CHK_STATUS_RET(GetNodeMemoryType(input_continuous_node, memory_type, "input"),
"[Get][MemType]fail for node:%s", input_continuous_node->GetName().c_str());
auto ret = AssignContinuousInputMemory(input_continuous_node, mem_clean_start, mem_clean_size, memory_type,
continuous_type, reverse_refresh);
if (ret != ge::SUCCESS) {
GELOGE(ret, "[Assign][Memory:Input:continuous]fail for node:%s", input_continuous_node->GetName().c_str());
return ret;
}
// Clean up atomic address, eg, hcom node
vector<int32_t> input_indexes;
// If GetListInt fail, input_indexes is empty.
(void)ge::AttrUtils::GetListInt(input_continuous_node->GetOpDesc(), ATOMIC_ATTR_INPUT_INDEX, input_indexes);
if (!input_indexes.empty() && input_indexes[0] == kAllInputAddrIsAtomic) {
// check whether there is an atomic conflict between the current node and the peer out node
if (!CheckInputIsSupportAtomic(input_continuous_node)) {
return ge::FAILED;
}
const auto &in_control_anchor = input_continuous_node->GetInControlAnchor();
GE_CHECK_NOTNULL(in_control_anchor);
for (const auto &peer_out_control_anchor : in_control_anchor->GetPeerOutControlAnchors()) {
GE_CHECK_NOTNULL(peer_out_control_anchor);
auto peer_out_node = peer_out_control_anchor->GetOwnerNode();
if (peer_out_node->GetType() == ATOMICADDRCLEAN) {
ret = SetAtomicCleanAttr(peer_out_node, {mem_clean_start}, {mem_clean_size}, memory_type);
if (ret != SUCCESS) {
GELOGE(ret, "[Set][AtomicCleanAttr]fail for node:%s", peer_out_node->GetName().c_str());
return ret;
}
}
}
}
return ge::SUCCESS;
}
Status GraphMemoryAssigner::AssignBufferPoolMemory() {
auto is_buffer_pool_mem_enable = [] (const ComputeGraphPtr &graph) -> bool {
for (NodePtr &node : graph->GetAllNodes()) {
auto op_desc = node->GetOpDesc();
if (op_desc == nullptr) {
continue;
}
bool has_attrs = op_desc->HasAttr(ATTR_NAME_BUFFER_POOL_ID) && op_desc->HasAttr(ATTR_NAME_BUFFER_POOL_SIZE);
if (has_attrs) {
return true;
}
}
return false;
};
auto root_graph = GraphUtils::FindRootGraph(compute_graph_);
GE_CHECK_NOTNULL(root_graph);
if (root_graph->GetGraphUnknownFlag()) {
GELOGI("[Check][Enable]Unknown root graph does not support buffer pool memory, graph:%s.",
compute_graph_->GetName().c_str());
return SUCCESS;
}
if (!is_buffer_pool_mem_enable(compute_graph_)) {
GELOGD("[Check][Enable]Buffer pool memory is not enable, graph:%s.", compute_graph_->GetName().c_str());
return SUCCESS;
}
map<int64_t, size_t> mem_type_to_offset;
for (const auto &pair : memory_offset_) {
mem_type_to_offset[pair.first] = pair.second.mem_offset_;
}
BufferPoolMemAssigner buffer_pool_mem_assigner(compute_graph_, mem_type_to_offset);
Status status = buffer_pool_mem_assigner.Assign();
if (status != SUCCESS) {
GELOGE(status, "[Assign][BufferPoolMem]Graph:%s.", compute_graph_->GetName().c_str());
REPORT_INNER_ERROR("E19999", "Failed to assign buffer pool memory, graph:%s.", compute_graph_->GetName().c_str());
return status;
}
int64_t mem_type = buffer_pool_mem_assigner.GetMemType();
auto iter = memory_offset_.find(mem_type);
if (iter == memory_offset_.end()) {
GELOGE(FAILED, "[Check][MemType]Memory type is not supported, graph:%s, mem type:%ld.",
compute_graph_->GetName().c_str(), mem_type);
REPORT_INNER_ERROR("E19999", "Memory type is not supported, graph:%s, mem type:%ld.",
compute_graph_->GetName().c_str(), mem_type);
return FAILED;
}
iter->second.mem_offset_ = buffer_pool_mem_assigner.GetMemOffset();
GELOGI("[Assign][BufferPoolMem]Assign buffer pool memory successfully, graph:%s, mem type:%ld, mem offset:%zu.",
compute_graph_->GetName().c_str(), mem_type, buffer_pool_mem_assigner.GetMemOffset());
return SUCCESS;
}
} // namespace ge
Reference in new issue View Git Blame Copy Permalink