TRT pool2d adaptive mode bugfix (#46802)

* draft with debug print

paddle/fluid/inference/tensorrt/plugin/pool_op_plugin.cu

@@ -321,16 +321,16 @@ int PoolPluginDynamic::enqueue(const nvinfer1::PluginTensorDesc *input_desc,
     paddings[1] = 0;
     output_shape[2] = 1;
     output_shape[3] = 1;
+    if (adaptive_) {
+      output_shape[2] = h;
+      output_shape[3] = w;
+    }
   } else {
     auto data_dim = CalcOutputSize(
         {h, w}, ceil_mode_, adaptive_, ksize_, strides_, paddings_);
     output_shape[2] = data_dim[0];
     output_shape[3] = data_dim[1];
   }
-  if (adaptive_) {
-    output_shape[2] = h;
-    output_shape[3] = w;
-  }
 
   if (pool_type_ == "max") {
     phi::funcs::MaxPool<float> pool_process;

paddle/phi/kernels/funcs/pooling.cu

@@ -460,7 +460,6 @@ void Pool2dDirectCUDAFunctor<PoolProcess, T>::operator()(
   const int stride_width = strides[1];
   const int padding_height = paddings[0];
   const int padding_width = paddings[1];
-
   int nthreads = batch_size * output_channels * output_height * output_width;
   auto pool_divmods =
       FastDivModForPooling(input_channels, output_width, output_height);
@@ -491,6 +490,7 @@ void Pool2dDirectCUDAFunctor<PoolProcess, T>::operator()(
                                        pool_compute,
                                        exclusive,
                                        output);
+
   } else {
     int thread_num = 1024;
 #ifdef WITH_NV_JETSON