[Dy2St] Optimize range_block_do performance (#69834)

Author: SigureMo

paddle/fluid/pybind/pir.cc

@@ -1690,30 +1690,27 @@ void BindInsertionPoint(pybind11::module *m) {
           return_value_policy::reference);
 }
 
-std::list<Operation *>::const_iterator list_offset(const Block *block,
-                                                   int start_idx) {
-  auto it = block->begin();
-  while (it != block->end() && start_idx--) ++it;
-  return it;
-}
-
 template <typename F, typename S>
 void range_block_do(const Block *block,
-                    std::vector<int> range,
+                    std::pair<size_t, size_t> range,
                     F fn,
                     S skip_fn) {
-  for (auto it = list_offset(block, range[0]);
-       it != list_offset(block, range[1]);
-       ++it) {
-    if (skip_fn(*it)) {
+  auto [start, end] = range;
+  if (start >= end) {
+    return;
+  }
+  auto it = block->begin();
+  std::advance(it, start);
+  for (size_t i = start; i < end && it != block->end(); ++i, ++it) {
+    if (skip_fn(it)) {
       continue;
     }
-    fn(*it);
+    fn(it);
   }
 }
 
 template <typename F>
-void range_block_do(const Block *block, std::vector<int> range, F fn) {
+void range_block_do(const Block *block, std::pair<size_t, size_t> range, F fn) {
   range_block_do(block, range, fn, [](Operation *op) { return false; });
 }
 
@@ -1754,8 +1751,8 @@ std::pair<std::vector<pir::Value>, std::unordered_set<pir::Value>>
 AnalysisMiddleVariable(const Program &program,
                        const std::vector<pir::Value> &forward_inputs,
                        const std::vector<pir::Value> &backward_outputs,
-                       const std::vector<int> &forward_range,
-                       const std::vector<int> &backward_range) {
+                       const std::pair<size_t, size_t> &forward_range,
+                       const std::pair<size_t, size_t> &backward_range) {
   std::vector<pir::Value> middle_values;
 
   std::unordered_set<pir::Value> backward_used_values;
@@ -1811,7 +1808,7 @@ using SplitedAttribute = std::map<std::string, std::vector<pir::Value>>;
 using SplitedResult = std::pair<SplitedProgram, SplitedAttribute>;
 
 static auto GetNoNeedBufferValue(const ::pir::Block *whole_block,
-                                 std::vector<int> range) {
+                                 std::pair<size_t, size_t> range) {
   // filter no need buffer values.
   std::unordered_set<::pir::Value> need_buffer_values;
   std::unordered_set<::pir::Value> no_need_buffer_values;
@@ -1926,8 +1923,8 @@ SplitedResult SplitForwardBackward(
     const std::vector<pir::Value> &forward_inputs_grads,
     const std::vector<pir::Value> &forward_params_grads,
     const std::vector<pir::Value> &forward_outputs_grads,
-    const std::vector<int> &forward_range,
-    const std::vector<int> &backward_range) {
+    const std::pair<size_t, size_t> &forward_range,
+    const std::pair<size_t, size_t> &backward_range) {
   std::vector<pir::Value> forward_in_out_values;
   for (auto &v :
        std::vector({&forward_inputs, &forward_outputs, &forward_params})) {

python/paddle/jit/dy2static/pir_partial_program.py

@@ -260,8 +260,8 @@ def split_forward_backward(self):
             self.x_grad_values,
             self.param_grad_values,
             self.out_grad_values,
-            list(self.forward_range),
-            list(self.backward_range),
+            self.forward_range,
+            self.backward_range,
         )
         return [fwd_prog, bwd_prog], prog_attr