fix runtime_context_cache bug when gpu model has an op runs only on cpu

test=develop

paddle/fluid/framework/operator.cc

@@ -1095,6 +1095,17 @@ Scope* OperatorWithKernel::PrepareData(
       if (!new_scope) {
         new_scope = &scope.NewScope();
       }
+      // For inference, if a gpu model has an op which could only run on CPU,
+      // each result of different input will be the same with the first one.
+      // The reason is that if a gpu tensor is the input of a cpu kernel,
+      // we will create a new cpu tensor in new scope.
+      // However, if enable_cache_runtime_context, we get the cpu tensor each
+      // time, not the gpu tensor.
+      // Thus, we set pre_scope_ = nullptr to trigger `new RuntimeContext()` in
+      // RunImpl().
+      if (enable_cache_runtime_context) {
+        pre_scope_ = nullptr;
+      }
 
       auto* trans_var = new_scope->Var(var_name);
       input_vars[i] = trans_var;

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -98,7 +98,10 @@ 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
+        // work on all fused ops.
         "expected_kernel_cache_pass",  //
+        "runtime_context_cache_pass"
   });
 
   use_gpu_ = true;
@@ -115,8 +118,7 @@ void GpuPassStrategy::EnableMkldnnQuantizer() {
 CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
   // NOTE the large fusions should be located in the front, so that they will
   // not be damaged by smaller ones.
-  passes_.assign({
-      "infer_clean_graph_pass",         //
+  passes_.assign({"infer_clean_graph_pass",         //
                   "attention_lstm_fuse_pass",       //
                   "seqconv_eltadd_relu_fuse_pass",  //
                   // "seqpool_concat_fuse_pass",    //
@@ -132,8 +134,10 @@ 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",  //
-  });
+                  "runtime_context_cache_pass"});
 
   use_gpu_ = false;
 }

paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc

@@ -110,11 +110,6 @@ void SetConfig(AnalysisConfig *cfg) {
   if (FLAGS_zero_copy) {
     cfg->SwitchUseFeedFetchOps(false);
   }
-  // Enable runtime_context_cache_pass, disabled by default since it doesn't
-  // cover all the cases.
-  // See detail: https://github.com/PaddlePaddle/Paddle/issues/16609
-  // https://github.com/PaddlePaddle/Paddle/issues/16841
-  cfg->pass_builder()->AppendPass("runtime_context_cache_pass");
 }
 
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {