复用runtime_op中的一部分内存

demos/yolo/yolo_test.cpp

@@ -86,7 +86,7 @@ void YoloDemo(const std::vector<std::string>& image_paths, const std::string& pa
 
   std::vector<std::shared_ptr<Tensor<float>>> outputs;
   graph.set_inputs("pnnx_input_0", inputs);
-  for (int i = 0; i < 1; ++i) {
+  for (int i = 0; i < 10000; ++i) {
     graph.Forward(true);
   }
   outputs = graph.get_outputs("pnnx_output_0");

include/runtime/runtime_op.hpp

@@ -53,6 +53,8 @@ struct RuntimeOperatorBase {
 
   int32_t end_forward_index = -1;
 
+  int32_t occur_forward_index = -1;
+
   /// Whether this operator has run in current execution
   bool has_forward = false;
 

source/runtime/runtime_ir.cpp

@@ -408,6 +408,7 @@ void RuntimeGraph::ReverseTopoSort() {
     } else {
       op->end_forward_index = last_forward_index;
     }
+    op->occur_forward_index = -1;
   }
 }
 

source/runtime/runtime_op.cpp

@@ -142,9 +142,14 @@ void RuntimeOperatorUtils<float>::InitOperatorOutput(
         }
 
         const auto& prev_runtime_op = operators.at(j);
-        if (!prev_runtime_op->output_operands || prev_runtime_op->end_forward_index == -1) {
+        if (!prev_runtime_op->output_operands || prev_runtime_op->occur_forward_index != -1) {
           continue;
         }
+
+        if (runtime_op->forward_index > prev_runtime_op->occur_forward_index) {
+          prev_runtime_op->occur_forward_index = -1;
+        }
+
         if (runtime_op->forward_index > prev_runtime_op->end_forward_index) {
           if (prev_runtime_op->output_operands->size() == operand_size) {
             has_found = true;
@@ -160,7 +165,7 @@ void RuntimeOperatorUtils<float>::InitOperatorOutput(
                   std::make_shared<ftensor>(output_tensor->raw_ptr(), output_tensor->shapes());
               CheckAndReshapeTensor(output_tensors->datas[b], operand_shapes);
             }
-            prev_runtime_op->end_forward_index = -1;
+            prev_runtime_op->occur_forward_index = runtime_op->end_forward_index;
           }
         }
       }