Merge pull request #7294 from jacquesqiao/add-back-priority

add back priority

paddle/framework/init.cc

@@ -72,7 +72,7 @@ bool InitDevices(const std::vector<std::string> &devices) {
     LOG(WARNING) << "Not specified CPU device, create CPU by Default.";
   }
   platform::DeviceContextPool::Init(places);
-  framework::UseALL();
+  // framework::UseALL();
   return true;
 }
 

paddle/framework/op_registry_test.cc

@@ -376,16 +376,16 @@ TEST(OperatorRegistrar, OpWithMultiKernel) {
   paddle::framework::UseCPU();
   op->Run(scope, cpu_place);
 
-  EXPECT_EQ(op_test_value, -20);
+  EXPECT_EQ(op_test_value, -9);
 
   // add cuda kernels
   paddle::framework::UseCUDA();
   op->Run(scope, cuda_place);
 
-  EXPECT_EQ(op_test_value, -30);
+  EXPECT_EQ(op_test_value, -10);
 
   // use cudnn kernel
   paddle::framework::UseCUDNN();
   op->Run(scope, cuda_place);
-  EXPECT_EQ(op_test_value, -40);
+  EXPECT_EQ(op_test_value, -20);
 }

paddle/framework/operator.cc

@@ -495,6 +495,22 @@ void OperatorWithKernel::Run(const Scope& scope,
   ExecutionContext ctx(*this, scope, *dev_ctx);
   auto expected_kernel_key = this->GetExpectedKernelType(ctx);
 
+  OpKernelMap& kernels = kernels_iter->second;
+
+  for (auto& candidate : kKernelPriority) {
+    auto candidate_key =
+        OpKernelType(expected_kernel_key.data_type_, std::get<0>(candidate),
+                     expected_kernel_key.data_layout_, std::get<1>(candidate));
+
+    if ((candidate_key == expected_kernel_key) ||
+        (kernels.count(candidate_key))) {
+      expected_kernel_key = candidate_key;
+      break;
+    }
+  }
+
+  VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
+
   Scope& new_scope = scope.NewScope();
 
   for (auto& var_name_item : this->Inputs()) {
@@ -525,10 +541,10 @@ void OperatorWithKernel::Run(const Scope& scope,
     }
   }
 
-  OpKernelMap& kernels = kernels_iter->second;
   auto kernel_iter = kernels.find(expected_kernel_key);
 
-  kernel_iter->second->Compute(ExecutionContext(*this, new_scope, *dev_ctx));
+  kernel_iter->second->Compute(ExecutionContext(
+      *this, new_scope, *pool.Get(expected_kernel_key.place_)));
 }
 
 proto::DataType OperatorWithKernel::IndicateDataType(

paddle/operators/fetch_op.cc

@@ -53,7 +53,7 @@ class FetchOp : public framework::OperatorBase {
     // FIXME(yuyang18): Should we assume the fetch operator always generate
     // CPU outputs?
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
-    auto &dev_ctx = *pool.Get(place);
+    auto &dev_ctx = *pool.Get(src_item.place());
 
     CopyFrom(src_item, platform::CPUPlace(), dev_ctx, &dst_item);
     dev_ctx.Wait();