[cherry-pick] c api update in PD_PredictorRun (#20705)

* improve the performance of capi in PD_PredictorRun (#20665)

* alter the capi of PD_PredictorRun to provide proper function, test=release/1.6

paddle/fluid/inference/capi/c_api.h

@@ -99,8 +99,8 @@ PADDLE_CAPI_EXPORT extern int* PD_GetPaddleTensorShape(const PD_Tensor* tensor,
 // AnalysisPredictor
 PADDLE_CAPI_EXPORT extern bool PD_PredictorRun(const PD_AnalysisConfig* config,
                                                PD_Tensor* inputs, int in_size,
-                                               PD_Tensor* output_data,
-                                               int** out_size, int batch_size);
+                                               PD_Tensor** output_data,
+                                               int* out_size, int batch_size);
 
 PADDLE_CAPI_EXPORT extern bool PD_PredictorZeroCopyRun(
     const PD_AnalysisConfig* config, PD_ZeroCopyData* inputs, int in_size,

paddle/fluid/inference/capi/pd_predictor.cc

@@ -26,10 +26,16 @@ using paddle::ConvertToACPrecision;
 extern "C" {
 
 bool PD_PredictorRun(const PD_AnalysisConfig* config, PD_Tensor* inputs,
-                     int in_size, PD_Tensor* output_data, int** out_size,
+                     int in_size, PD_Tensor** output_data, int* out_size,
                      int batch_size) {
   PADDLE_ENFORCE_NOT_NULL(config);
-  auto predictor = paddle::CreatePaddlePredictor(config->config);
+  static std::map<std::string, std::unique_ptr<paddle::PaddlePredictor>>
+      predictors;
+  if (!predictors.count(config->config.model_dir())) {
+    predictors[config->config.model_dir()] =
+        paddle::CreatePaddlePredictor(config->config);
+  }
+  auto& predictor = predictors[config->config.model_dir()];
   std::vector<paddle::PaddleTensor> in;
   for (int i = 0; i < in_size; ++i) {
     in.emplace_back(inputs->tensor);
@@ -37,10 +43,11 @@ bool PD_PredictorRun(const PD_AnalysisConfig* config, PD_Tensor* inputs,
   std::vector<paddle::PaddleTensor> out;
   if (predictor->Run(in, &out, batch_size)) {
     int osize = out.size();
+    *output_data = new PD_Tensor[osize];
     for (int i = 0; i < osize; ++i) {
-      output_data[i].tensor = out[i];
+      output_data[i]->tensor = out[i];
     }
-    *out_size = &osize;
+    *out_size = osize;
     return true;
   }
   return false;

paddle/fluid/inference/tests/api/analyzer_capi_pd_tensor_tester.cc

@@ -21,6 +21,7 @@ limitations under the License. */
 #include <string>
 #include <vector>
 #include "paddle/fluid/inference/capi/c_api.h"
+#include "paddle/fluid/inference/capi/c_api_internal.h"
 #include "paddle/fluid/inference/tests/api/tester_helper.h"
 
 namespace paddle {
@@ -56,16 +57,15 @@ void PD_run() {
   PD_SetPaddleTensorData(input, buf);
 
   PD_Tensor* out_data = PD_NewPaddleTensor();
-  int* out_size;
-  PD_PredictorRun(config, input, 1, out_data, &out_size, 1);
-  LOG(INFO) << *out_size;
+  int out_size;
+  PD_PredictorRun(config, input, 1, &out_data, &out_size, 1);
+  LOG(INFO) << out_size;
   LOG(INFO) << PD_GetPaddleTensorName(out_data);
   LOG(INFO) << PD_GetPaddleTensorDType(out_data);
   PD_PaddleBuf* b = PD_GetPaddleTensorData(out_data);
-  LOG(INFO) << PD_PaddleBufLength(b);
+  LOG(INFO) << PD_PaddleBufLength(b) / sizeof(float);
   float* result = static_cast<float*>(PD_PaddleBufData(b));
   LOG(INFO) << *result;
-  PD_PaddleBufResize(b, 500);
   PD_DeletePaddleTensor(input);
   int* size;
   PD_GetPaddleTensorShape(out_data, &size);
@@ -132,16 +132,15 @@ void buffer_run() {
   PD_SetPaddleTensorData(input, buf);
 
   PD_Tensor* out_data = PD_NewPaddleTensor();
-  int* out_size;
-  PD_PredictorRun(config, input, 1, out_data, &out_size, 1);
-  LOG(INFO) << *out_size;
+  int out_size;
+  PD_PredictorRun(config, input, 1, &out_data, &out_size, 1);
+  LOG(INFO) << out_size;
   LOG(INFO) << PD_GetPaddleTensorName(out_data);
   LOG(INFO) << PD_GetPaddleTensorDType(out_data);
   PD_PaddleBuf* b = PD_GetPaddleTensorData(out_data);
-  LOG(INFO) << PD_PaddleBufLength(b);
+  LOG(INFO) << PD_PaddleBufLength(b) / sizeof(float);
   float* result = static_cast<float*>(PD_PaddleBufData(b));
   LOG(INFO) << *result;
-  PD_PaddleBufResize(b, 500);
   PD_DeletePaddleTensor(input);
   PD_DeletePaddleBuf(buf);
 }