Exhaustive search for cuDNN conv. (#14286)

* exhaustive search for cuDNN conv.
* Refine code and add unit testing.
* Fix model load in fluid/inference and unit testing in conv2d
* Follow comments.
* Fix compiling test=develop

Author: qingqing01

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <algorithm>
 #include <array>
 #include <string>
 #include <vector>

paddle/fluid/inference/api/analysis_predictor.h

@@ -13,6 +13,8 @@
 // limitations under the License.
 
 #pragma once
+#include <algorithm>
+#include <map>
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/naive_executor.h"

paddle/fluid/inference/api/api.cc

@@ -16,7 +16,6 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/platform/enforce.h"
-#include "paddle_inference_api.h"
 
 namespace paddle {
 

paddle/fluid/inference/api/helper.h

@@ -16,13 +16,14 @@
 
 #include <glog/logging.h>
 #include <sys/time.h>
+#include <algorithm>
 #include <chrono>  // NOLINT
 #include <numeric>
 #include <sstream>
 #include <string>
 #include <vector>
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/string/printf.h"
-#include "paddle_inference_api.h"
 
 namespace paddle {
 namespace inference {

paddle/fluid/inference/io.cc

@@ -59,7 +59,8 @@ void ReadBinaryFile(const std::string& filename, std::string* contents) {
 bool IsPersistable(const framework::VarDesc* var) {
   if (var->Persistable() &&
       var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
-      var->GetType() != framework::proto::VarType::FETCH_LIST) {
+      var->GetType() != framework::proto::VarType::FETCH_LIST &&
+      var->GetType() != framework::proto::VarType::RAW) {
     return true;
   }
   return false;

paddle/fluid/inference/tensorrt/engine.h

@@ -134,7 +134,7 @@ class TensorRTEngine : public EngineBase {
   std::unordered_map<std::string /*name*/, std::unique_ptr<framework::Tensor>>
       weight_map;
 
-  // TODO: (NHZLX)
+  // TODO(NHZLX)
   // In the normal case, the paddle-trt exists bug when runing the googlenet.
   // When there are more than two convolutions of 1 * 1 with the same input, the
   // paddle-tensorrt will do the merging optimization, which fuse those conv

paddle/fluid/operators/add_position_encoding_op.h

@@ -66,9 +66,10 @@ class AddPositionEncodingKernel : public framework::OpKernel<T> {
           x_lod.empty() ? max_seq_len : x_lod[0][i + 1] - x_lod[0][i];
       for (int j = 0; j < max_length; ++j) {
         for (int k = 0; k < half_size; ++k) {
-          const double val = (half_size > 1)
-                                 ? j / pow(10000.0, double(k) / (half_size - 1))
-                                 : j / 10000.0;
+          const double val =
+              (half_size > 1)
+                  ? j / pow(10000.0, static_cast<double>(k) / (half_size - 1))
+                  : j / 10000.0;
           dst_ptr[k] = src_ptr[k] * alpha + sin(val) * beta;
           dst_ptr[half_size + k] =
               src_ptr[half_size + k] * alpha + cos(val) * beta;