add expected_kernel_cache_pass

test=develop

paddle/fluid/framework/ir/CMakeLists.txt

@@ -71,6 +71,7 @@ pass_library(transpose_flatten_concat_fuse_pass inference)
 pass_library(identity_scale_op_clean_pass base)
 pass_library(sync_batch_norm_pass base)
 pass_library(runtime_context_cache_pass base)
+pass_library(expected_kernel_cache_pass base)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will

paddle/fluid/framework/ir/expected_kernel_cache_pass.cc

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/ir/expected_kernel_cache_pass.h"
+#include <memory>
+#include "paddle/fluid/framework/operator.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> ExpectedKernelCachePass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  VLOG(3) << "Applies Expected Kernel Cache strategy.";
+  for (const Node* n : graph->Nodes()) {
+    if (n->IsOp()) {
+      n->Op()->SetAttr(kEnableCacheExpectedKernel, true);
+    }
+  }
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(expected_kernel_cache_pass,
+              paddle::framework::ir::ExpectedKernelCachePass);

paddle/fluid/framework/ir/expected_kernel_cache_pass.h

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <memory>
+#include "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class ExpectedKernelCachePass : public Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/operator.cc

@@ -895,7 +895,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
-  if (!kernel_type_) {
+  if (!HasAttr(kEnableCacheExpectedKernel) || !kernel_type_) {
     ChooseKernel(*runtime_ctx, scope, place);
   }
 

paddle/fluid/framework/operator.h

@@ -70,6 +70,12 @@ constexpr char kNewGradSuffix[] = "@NEWGRAD@";
 /// this Op's execution to save the elapsed time.
 constexpr char kEnableCacheRuntimeContext[] = "@ENABLE_CACHE_RUNTIME_CONTEXT@";
 
+/// If an Op has attribtue kEnableCacheExpectedKernel, it means that in a same
+/// name scope and same place, since the expected kerenl of this Op does not
+/// change in the execution, it could be recorded only at the first iteration of
+/// this Op's execution to save the elapsed time.
+constexpr char kEnableCacheExpectedKernel[] = "@ENABLE_CACHE_EXPECTED_KERNEL@";
+
 /// If an Op has this attribute, all its kernels should calculate output
 /// variable's shape in the corresponding Compute() function. And
 /// OperatorWithKernel::RunImpl() would skip call this Op's InferShape()

paddle/fluid/inference/api/analysis_config.cc

@@ -203,6 +203,7 @@ void AnalysisConfig::Update() {
       pass_builder()->InsertPass(3, "tensorrt_subgraph_pass");
     }
     pass_builder()->DeletePass("runtime_context_cache_pass");
+    pass_builder()->DeletePass("expected_kernel_cache_pass");
   }
 
   if (use_mkldnn_) {

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -81,6 +81,7 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
         "conv_elementwise_add2_act_fuse_pass",  //
         "conv_elementwise_add_fuse_pass",       //
         "runtime_context_cache_pass",           //
+        "expected_kernel_cache_pass",           //
 #endif
   });
 
@@ -117,6 +118,7 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
       "is_test_pass",                  //
       "identity_scale_op_clean_pass",  //
       "runtime_context_cache_pass",    //
+      "expected_kernel_cache_pass",    //
   });
   use_gpu_ = false;
 }