remove unused expected_kernel_cache_pass (#17486)

test=develop

Author: luotao1

paddle/fluid/framework/details/build_strategy.cc

@@ -171,11 +171,6 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
       AppendPass("runtime_context_cache_pass");
     }
 
-    if (strategy_.cache_expected_kernel_) {
-      VLOG(10) << "Add expected_kernel_cache_pass";
-      AppendPass("expected_kernel_cache_pass");
-    }
-
     AppendMultiDevPass(strategy_);
 
     if (strategy_.fuse_all_reduce_ops_) {
@@ -371,7 +366,6 @@ USE_PASS(fuse_sgd_op_pass);
 USE_PASS(fuse_momentum_op_pass);
 USE_PASS(fuse_all_reduce_op_pass);
 USE_PASS(runtime_context_cache_pass);
-USE_PASS(expected_kernel_cache_pass);
 USE_PASS(record_skip_memory_opt_vars_pass);
 #ifdef PADDLE_WITH_MKLDNN
 USE_PASS(mkldnn_placement_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -109,7 +109,6 @@ struct BuildStrategy {
   bool remove_unnecessary_lock_{true};
 
   bool cache_runtime_context_{false};
-  bool cache_expected_kernel_{true};
   std::unordered_set<std::string> mkldnn_enabled_op_types_;
 
   // NOTE:

paddle/fluid/framework/ir/CMakeLists.txt

@@ -72,7 +72,6 @@ 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)
 pass_library(quant_conv2d_dequant_fuse_pass inference)
 pass_library(fillconstant_elementwisemul_fuse inference)
 pass_library(shuffle_channel_detect_pass inference)

paddle/fluid/framework/ir/expected_kernel_cache_pass.cc

@@ -71,12 +71,6 @@ 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/framework/ir/expected_kernel_cache_pass.h

@@ -235,7 +235,6 @@ 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/framework/operator.h

@@ -98,9 +98,8 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
         "conv_elementwise_add_fuse_pass",       //
 #endif                                          //
         "transpose_flatten_concat_fuse_pass",
-        // following two passes should be located in the last, since they will
+        // following pass should be located in the last, since it will
         // work on all fused ops.
-        "expected_kernel_cache_pass",  //
         "runtime_context_cache_pass"
   });
 
@@ -134,9 +133,8 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
                   "conv_bn_fuse_pass",             //
                   "conv_eltwiseadd_bn_fuse_pass",  //
                   "is_test_pass",                  //
-                  // following two passes should be located in the last, since
-                  // they will work on all fused ops.
-                  "expected_kernel_cache_pass",  //
+                  // following pass should be located in the last, since
+                  // it will work on all fused ops.
                   "runtime_context_cache_pass"});
 
   use_gpu_ = false;

paddle/fluid/inference/api/analysis_config.cc

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <mutex>  // NOLINT // for call_once
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 
@@ -1492,10 +1493,6 @@ All parameter, weight, gradient are variables in Paddle.
           "cache_runtime_context",
           [](const BuildStrategy &self) { return self.cache_runtime_context_; },
           [](BuildStrategy &self, bool b) { self.cache_runtime_context_ = b; })
-      .def_property(
-          "cache_expected_kernel",
-          [](const BuildStrategy &self) { return self.cache_expected_kernel_; },
-          [](BuildStrategy &self, bool b) { self.cache_expected_kernel_ = b; })
       .def_property(
           "mkldnn_enabled_op_types",
           [](const BuildStrategy &self) {