Add FLAGS_force_sync_ops for executor (#68467)

From00 · web-flow · commit 2aebcd88c95a · 2024-10-09T20:28:44.000+08:00
* Add sync_op_after_launch config for executor

* Update code

* Update code

* Update code
diff --git a/paddle/fluid/framework/new_executor/instruction/instruction_base.cc b/paddle/fluid/framework/new_executor/instruction/instruction_base.cc
@@ -197,8 +197,6 @@ InstructionBase::InstructionBase(size_t id, const phi::Place& place)
       no_need_buffer_values_() {
   id_ = id;
 
-  is_artificial_ = false;
-
   if (phi::is_cpu_place(place)) {
     type_ = OpFuncType::kCpuSync;
   } else {
diff --git a/paddle/fluid/framework/new_executor/instruction/instruction_base.h b/paddle/fluid/framework/new_executor/instruction/instruction_base.h
@@ -43,6 +43,9 @@ class InstructionBase {
   bool IsArtificial() const { return is_artificial_; }
   void SetArtificial(bool is_artificial) { is_artificial_ = is_artificial; }
 
+  bool IsSyncAfterLaunch() const { return sync_after_launch_; }
+  void SetSyncAfterLaunch(bool sync) { sync_after_launch_ = sync; }
+
   OpFuncType KernelType() const;
   void SetKernelType(OpFuncType type) { type_ = type; }
 
@@ -176,8 +179,12 @@ class InstructionBase {
  protected:
   size_t id_;
 
-  bool is_artificial_;  // Instruction is artificial means that it is only used
-                        // to assist scheduling and no need to be executed.
+  bool is_artificial_{
+      false};  // Instruction is artificial means that it is only used
+               // to assist scheduling and no need to be executed.
+
+  bool sync_after_launch_{false};
+
   OpFuncType type_;
 
   // dist attrs：lower value, higher priority
diff --git a/paddle/fluid/framework/new_executor/interpreter/execution_config.cc b/paddle/fluid/framework/new_executor/interpreter/execution_config.cc
@@ -17,10 +17,25 @@
 #include <set>
 #include <thread>
 
+#include "paddle/common/flags.h"
 #include "paddle/fluid/platform/device/ipu/ipu_info.h"
 #include "paddle/phi/backends/device_manager.h"
 #include "paddle/phi/backends/gpu/gpu_info.h"
 #include "paddle/phi/backends/xpu/xpu_info.h"
+#include "paddle/utils/string/string_helper.h"
+
+// FLAGS_force_sync_ops is used to finer control the op-sync in executor.
+// The format is: "micro_batch_id, job_name, op_id, op_name | micro_batch_id,
+// job_name, op_id, op_name | ...". Keep spaces to syncs all name/id. Example:
+// 1. sync the recv_v2 op in the second backward-job of 1F1B scheduling:
+// FLAGS_force_sync_ops="1, backward, , recv_v2"
+// 2. sync the full op with op_id=5: FLAGS_force_sync_ops=" , , 5, full"
+// 3. sync all ops in the first default-job: FLAGS_force_sync_ops="0,default,,
+// 4. sync all ops in the forward-job and backward-job: FLAGS_force_sync_ops=" ,
+// forward, , | , backward, , , "
+PHI_DEFINE_EXPORTED_string(force_sync_ops,
+                           "",
+                           "Pattern to force sync ops in executor.");
 
 PD_DECLARE_bool(new_executor_serial_run);
 
@@ -149,4 +164,65 @@ void ExecutionConfig::Log(int log_level) {
   VLOG(log_level) << log_str.str();
 }
 
+std::set<std::pair<int, std::string>> GetForceSyncOps(
+    int micro_batch_id, const std::string& job_name) {
+  std::set<std::pair<int, std::string>> force_sync_ops;
+  std::stringstream ss(paddle::string::erase_spaces(FLAGS_force_sync_ops));
+  std::string item;
+
+  while (std::getline(ss, item, '|')) {
+    item += ",";  // The comma at the end of the string will be ignored in
+                  // std::getline
+    std::stringstream item_stream(item);
+    std::vector<std::string> tokens;
+    std::string token;
+    while (std::getline(item_stream, token, ',')) {
+      VLOG(1) << "token: " << token;
+      tokens.push_back(token);
+    }
+
+    PADDLE_ENFORCE_EQ(
+        tokens.size(),
+        4,
+        phi::errors::InvalidArgument("Invalid force_sync_ops format: \"%s\", "
+                                     "FLAGS_force_sync_ops=\"%s\"",
+                                     item,
+                                     FLAGS_force_sync_ops));
+
+    int micro_batch_id_;
+    if (tokens[0] == "") {
+      micro_batch_id_ = -1;
+    } else {
+      micro_batch_id_ = std::stoi(tokens[0]);
+    }
+    if (micro_batch_id_ != micro_batch_id && micro_batch_id_ != -1) {
+      continue;
+    }
+
+    if (tokens[1] != job_name && tokens[1] != "") {
+      continue;
+    }
+
+    int op_id;
+    if (tokens[2] == "") {
+      op_id = -1;
+    } else {
+      op_id = std::stoi(tokens[2]);
+    }
+    std::string op_name = tokens[3];
+    force_sync_ops.insert({op_id, op_name});
+  }
+
+  if (!force_sync_ops.empty()) {
+    std::stringstream ss;
+    ss << "job_name: " << job_name << ", micro_batch_id: " << micro_batch_id
+       << ", force_sync_ops: ";
+    for (auto& pair : force_sync_ops) {
+      ss << "(" << pair.first << ", " << pair.second << ") ";
+    }
+    VLOG(6) << ss.str();
+  }
+  return force_sync_ops;
+}
+
 }  // namespace paddle::framework::interpreter
diff --git a/paddle/fluid/framework/new_executor/interpreter/execution_config.h b/paddle/fluid/framework/new_executor/interpreter/execution_config.h
@@ -26,7 +26,6 @@ namespace interpreter {
 
 struct ExecutionConfig {
   bool create_local_scope{true};
-
   bool used_for_cinn{false};
   bool used_for_control_flow_op{false};
   bool used_for_jit{false};
@@ -35,6 +34,10 @@ struct ExecutionConfig {
   size_t device_num_threads{0};
   size_t host_num_threads{0};
 
+  std::set<std::pair<int, std::string>>
+      force_sync_ops;  // set{pair<op_id, name>}, -1 matches any op_id, ""
+                       // matches any name
+
   std::set<std::string> force_root_scope_vars;
   std::set<std::string> jit_input_vars;
   std::set<std::string> skip_gc_vars;
@@ -43,6 +46,9 @@ struct ExecutionConfig {
   void Log(int log_level);
 };
 
+std::set<std::pair<int, std::string>> GetForceSyncOps(
+    int micro_batch_id, const std::string& job_name);
+
 }  // namespace interpreter
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/framework/new_executor/pir_interpreter.cc b/paddle/fluid/framework/new_executor/pir_interpreter.cc
@@ -765,6 +765,33 @@ void PirInterpreter::AnalyseExecuteOrderForTrace(
   }
 }
 
+void PirInterpreter::AnalyzeForceSyncOps() {
+  for (auto& ins : vec_instruction_base_) {
+    ins->SetSyncAfterLaunch(FLAGS_benchmark);
+
+    // Analyze force sync op set by FLAGS_force_sync_op
+    int op_id = ins->Id();
+    std::string op_name = ins->Name();
+    std::string unused_prefix = "pd_op.";
+    auto pos = op_name.find(unused_prefix);
+    if (pos != std::string::npos) {
+      op_name.erase(pos, unused_prefix.size());
+    }
+
+    for (auto& pair : execution_config_.force_sync_ops) {
+      int sync_op_id = pair.first;
+      std::string sync_op_name = pair.second;
+      if ((sync_op_id == op_id || sync_op_id == -1) &&
+          (sync_op_name == op_name || sync_op_name == "")) {
+        VLOG(8) << "Force sync op: "
+                << "sync_op_id=" << sync_op_id << ", op_id=" << op_id
+                << ", sync_op_name=" << sync_op_name << ", op_name=" << op_name;
+        ins->SetSyncAfterLaunch(true);
+      }
+    }
+  }
+}
+
 void PirInterpreter::BuildInstruction() {
   VLOG(6) << "Build Instructions for pir ... ";
   vec_instruction_base_.clear();
@@ -1900,7 +1927,7 @@ void PirInterpreter::RunInstructionBase(InstructionBase* instr_node) {
         instr_node->Run();
       }
 
-      if (FLAGS_benchmark) {
+      if (instr_node->IsSyncAfterLaunch()) {
         instr_node->DeviceContext().Wait();
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
         PADDLE_ENFORCE_GPU_SUCCESS(platform::GpuGetLastError());
@@ -2003,6 +2030,9 @@ void PirInterpreter::PreAnalysis() {
                               ir_instruction_scheduling_priority_less);
   VLOG(4) << "Done AnalyseExecuteOrderForTrace";
 
+  AnalyzeForceSyncOps();
+  VLOG(4) << "Done AnalyzeForceSyncOps";
+
   UpdateSyncOpNum();
   VLOG(4) << "Done UpdateSyncOpNum";
 
diff --git a/paddle/fluid/framework/new_executor/pir_interpreter.h b/paddle/fluid/framework/new_executor/pir_interpreter.h
@@ -129,9 +129,11 @@ class PirInterpreter : public InterpreterBaseImpl {
   void UpdateSyncOpNum();
   void UpdateNcclOpNum();
   void UpdateOneDNNOpNum();
+
   void AnalyseExecuteOrderForTrace(
       std::map<size_t, std::set<size_t>> op_downstream_map,
       InstructionSchedulingPriorityLess compare);
+  void AnalyzeForceSyncOps();
   void ConstructEventForJitInput();
   void CalculateLastLiveOps();
 
diff --git a/paddle/fluid/framework/new_executor/standalone_executor.cc b/paddle/fluid/framework/new_executor/standalone_executor.cc
@@ -87,6 +87,8 @@ StandaloneExecutor::StandaloneExecutor(const phi::Place& place,
     interpreter::ExecutionConfig execution_config;
     execution_config.create_local_scope = false;
     execution_config.skip_gc_vars = job->SkipGcVars();
+    execution_config.force_sync_ops =
+        interpreter::GetForceSyncOps(micro_batch_id, job_type);
 
     // TODO(phlrain) we only support cpu for now
     if (FLAGS_enable_pir_in_executor) {
