remove slowing down pass (#43750)

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -20,6 +20,7 @@
 #include <miopen/miopen.h>
 #endif
 #include <glog/logging.h>
+#include <algorithm>
 #include <sstream>
 
 namespace paddle {
@@ -60,6 +61,12 @@ void PaddlePassBuilder::DeletePass(const std::string &pass_type) {
   }
 }
 
+size_t PaddlePassBuilder::GetPassIndex(const std::string &pass_type) {
+  auto iter = std::find(std::begin(passes_), std::end(passes_), pass_type);
+  if (iter == std::end(passes_)) return -1;
+  return std::distance(std::begin(passes_), iter);
+}
+
 void PaddlePassBuilder::InsertPass(size_t idx, const std::string &pass_type) {
   passes_.insert(std::begin(passes_) + idx, pass_type);
 }
@@ -268,7 +275,11 @@ void CpuPassStrategy::EnableMKLDNN() {
 #ifdef PADDLE_WITH_MKLDNN
   if (!use_mkldnn_) {
     passes_.insert(passes_.begin(), "mkldnn_placement_pass");
-
+    int id = GetPassIndex("gpu_cpu_reshape2_matmul_fuse_pass");
+    // this pass slows down FC mkldnn int8 operator
+    if (id != -1) {
+      passes_.erase(passes_.begin() + id);
+    }
     for (auto &pass : std::vector<std::string>({
              "depthwise_conv_mkldnn_pass",    //
              "conv_bn_fuse_pass",             // Execute BN passes again to

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -75,6 +75,10 @@ class PD_INFER_DECL PaddlePassBuilder {
   /// \param[in] pass_type the certain pass type to be deleted.
   void DeletePass(const std::string &pass_type);
 
+  /// \brief Get the certain position of a pass.
+  //  \param[in] pass_type the type of insert pass.
+  size_t GetPassIndex(const std::string &pass_type);
+
   /// \brief Delete all the passes.
   void ClearPasses();
 

python/paddle/fluid/contrib/slim/quantization/quant2_int8_mkldnn_pass.py

@@ -411,7 +411,7 @@ class Quant2Int8MkldnnPass(object):
         graph = self._apply_pass(graph, 'multi_gru_seq_fuse_pass')
         graph = self._apply_pass(graph, 'seq_concat_fc_fuse_pass')
         graph = self._apply_pass(graph, 'gpu_cpu_squeeze2_matmul_fuse_pass')
-        graph = self._apply_pass(graph, 'gpu_cpu_reshape2_matmul_fuse_pass')
+        #graph = self._apply_pass(graph, 'gpu_cpu_reshape2_matmul_fuse_pass')
         graph = self._apply_pass(graph, 'gpu_cpu_flatten2_matmul_fuse_pass')
         graph = self._apply_pass(graph, 'matmul_v2_scale_fuse_pass')
         graph = self._apply_pass(graph, 'squared_mat_sub_fuse_pass')