From eb32746a43f3c8c66f61c0ab58f191bd5796c6f3 Mon Sep 17 00:00:00 2001
From: Wang Bojun <105858416+wwbitejotunn@users.noreply.github.com>
Date: Fri, 14 Oct 2022 14:37:35 +0800
Subject: [PATCH] TRT pool2d adaptive mode bugfix (#46802)

* draft with debug print
---
 paddle/fluid/inference/tensorrt/plugin/pool_op_plugin.cu | 8 ++++----
 paddle/phi/kernels/funcs/pooling.cu                      | 2 +-
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/paddle/fluid/inference/tensorrt/plugin/pool_op_plugin.cu b/paddle/fluid/inference/tensorrt/plugin/pool_op_plugin.cu
index 21eb89d135..52959d8f90 100644
--- a/paddle/fluid/inference/tensorrt/plugin/pool_op_plugin.cu
+++ b/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;
diff --git a/paddle/phi/kernels/funcs/pooling.cu b/paddle/phi/kernels/funcs/pooling.cu
index 875fa92002..a1bb927b73 100644
--- a/paddle/phi/kernels/funcs/pooling.cu
+++ b/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
-- 
GitLab