[IR] Support GC and TraceRun for NewIr InterpreterCore #55772
Conversation
… dev/dependency_new_ir_exe
… dev/construct_event
… dev/construct_event
… dev/construct_event
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
❌ The PR is not created using PR's template. You can refer to this Demo. |
zhangbo9674
changed the title
| @@ -37,6 +37,19 @@ InterpreterCoreEventGarbageCollector::InterpreterCoreEventGarbageCollector( | |||
| } | |||
| } | |||
|
|
|||
| InterpreterCoreEventGarbageCollector::InterpreterCoreEventGarbageCollector( | |||
| const std::vector<std::unique_ptr<InstructionBase>>& vec_instruction) { | |||
There was a problem hiding this comment.
后续我们讨论下参数命名规范，是否需要以 type 的前缀来标记，比如vec_、map_。目前有的采用instructions 这种带 s后缀来表示 list信息的（我个人是倾向于后者的，因为可以跟类型解耦。）
| /*allow_spinning*/ true, | ||
| /*track_task*/ false); | ||
| queue_ = CreateSingleThreadedWorkQueue(options); | ||
| for (auto& instruc : vec_instruction) { |
There was a problem hiding this comment.
对于for-loop的缩写变量命令也可以留意一下，instr比 instruc更常见？或者不想起名，直接instruction来用（搭配前面的 instructions）可以很好区别。
| } | ||
| return -1; |
There was a problem hiding this comment.
这里为什么不直接PADDLE_THROW？caller需要处理这个-1么？
|
|
||
| auto op_downstream_map = dependency_builder_.OpDownstreamMap(); | ||
|
|
||
| VLOG(4) << "1"; |
|
|
||
| std::stringstream ss; | ||
| ss << "trace order: "; | ||
| for (size_t idx = 0; idx < trace_execute_order_.size(); idx++) { |
There was a problem hiding this comment.
这里总会执行的，建议使用 if(VLOG_IS_ON(6)){XXX}
| @@ -1725,5 +1686,455 @@ void NewIRInterpreter::BuildInstructionDependences() { | |||
| } | |||
| } | |||
|
|
|||
| void NewIRInterpreter::RecordMemcpyD2H(InstructionBase* instr_node) { | |||
| // NOTE(zhiqiu): hot fix for jit input var | |||
| if (instr_node->Name() == "pd.memcpy_d2h") { | |||
There was a problem hiding this comment.
后续我们要不要维护一个新IR下的头文件，定义所有这些特殊Op name，这样我们就知道初期我们累计了多少TODO的分支逻辑了，后续也好专项清理
| VLOG(4) << "var_ref_count_.size() : " << var_ref_count_.size(); | ||
| for (size_t i = 0; i < last_live_ops_.size(); ++i) { | ||
| std::set<size_t> minumum_last_live_ops; | ||
| for (auto val : last_live_ops_[i]) { |
There was a problem hiding this comment.
这个是LOG相关的for循环？看起来总会执行？
| refs_.emplace_back(std::make_shared<interpreter::VarRefInfo>( | ||
| var_ref_count_[i], variable_list_[i])); | ||
| } | ||
| VLOG(4) << "calculate last_live_ops_ 3"; |
PR types
New features
PR changes
Others
Description
基于新 IR InstructionBase 体系，支持 Gc 及 TraceRun，跑通 standalone_executor_new_ir_test 单测。
后续工作：
Others
Pcard-67164