From 5036cf03872a1a1b68cd974e21193ae82f5da071 Mon Sep 17 00:00:00 2001
From: QI JUN <qijun1994@hotmail.com>
Date: Fri, 29 Dec 2017 16:43:10 +0800
Subject: [PATCH] add helper function to get appropriate DeviceContext (#7066)

* add helper function to get appropriate DeviceContext
---
 paddle/framework/data_transform.h       |  5 ++--
 paddle/framework/data_transform_test.cc | 15 ++++++------
 paddle/framework/operator.cc            | 32 ++++++++++++++++++-------
 3 files changed, 33 insertions(+), 19 deletions(-)

diff --git a/paddle/framework/data_transform.h b/paddle/framework/data_transform.h
index 2191dd3783..bd6d301c12 100644
--- a/paddle/framework/data_transform.h
+++ b/paddle/framework/data_transform.h
@@ -27,9 +27,8 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-using DataTransformFn =
-    std::function<void(const std::vector<platform::DeviceContext*> ctx,
-                       const Variable& in, Variable* out)>;
+using DataTransformFn = std::function<void(const platform::DeviceContext* ctx,
+                                           const Variable& in, Variable* out)>;
 using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
 
 struct KernelTypePairHash {
diff --git a/paddle/framework/data_transform_test.cc b/paddle/framework/data_transform_test.cc
index 4e2141ecd2..5f05e881fa 100644
--- a/paddle/framework/data_transform_test.cc
+++ b/paddle/framework/data_transform_test.cc
@@ -54,18 +54,18 @@ auto kernel1 = GenFromBit({0, 0, 0, 1});
 auto kernel2 = GenFromBit({0, 0, 1, 0});
 auto kernel3 = GenFromBit({0, 0, 1, 1});
 
-void TransDataType_t(std::vector<platform::DeviceContext*> ctx,
-                     const Variable& in, Variable* out) {
+void TransDataType_t(const platform::DeviceContext* ctx, const Variable& in,
+                     Variable* out) {
   test_value++;
 }
 
-void TransDataLayout_t(std::vector<platform::DeviceContext*> ctx,
-                       const Variable& in, Variable* out) {
+void TransDataLayout_t(const platform::DeviceContext* ctx, const Variable& in,
+                       Variable* out) {
   test_value--;
 }
 
-void TransLibraryType_t(std::vector<platform::DeviceContext*> ctx,
-                        const Variable& in, Variable* out) {
+void TransLibraryType_t(const platform::DeviceContext* ctx, const Variable& in,
+                        Variable* out) {
   test_value += 2;
 }
 
@@ -83,7 +83,8 @@ TEST(DataTransform, Register) {
   using namespace paddle::platform;
 
   auto& instance = DataTransformFnMap::Instance();
-  std::vector<DeviceContext*> ctx;
+  ASSERT_EQ(instance.Map().size(), 3UL);
+  DeviceContext* ctx = nullptr;
   paddle::framework::Variable in;
   paddle::framework::Variable out;
 
diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index c0be11294c..a3ce96c409 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -384,6 +384,24 @@ class RuntimeInferShapeContext : public InferShapeContext {
   const Scope& scope_;
 };
 
+const platform::DeviceContext* GetDeviceContext(
+    framework::KernelTypePair& kernel_pair) {
+  auto& actual_kernel_key = kernel_pair.first;
+  auto& expected_kernel_key = kernel_pair.second;
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+
+  if (platform::is_gpu_place(actual_kernel_key.place_) &&
+      platform::is_cpu_place(expected_kernel_key.place_)) {
+    return pool.Get(actual_kernel_key.place_);
+  } else if (platform::is_cpu_place(actual_kernel_key.place_) &&
+             platform::is_gpu_place(expected_kernel_key.place_)) {
+    return pool.Get(expected_kernel_key.place_);
+  } else {
+    PADDLE_THROW(
+        "Currently, model parallelism is only supported between CPU and CUDA");
+  }
+}
+
 void OperatorWithKernel::Run(const Scope& scope,
                              const platform::Place& place) const {
   RuntimeInferShapeContext infer_shape_ctx(*this, scope);
@@ -418,9 +436,9 @@ void OperatorWithKernel::Run(const Scope& scope,
                       "CPU and other devices. For example, multi-GPU model "
                       "parallelism will failed.");
   } else {
+    auto kernel_pair = std::make_pair(actual_kernel_key, expected_kernel_key);
     const DataTransformFn* trans_fun =
-        DataTransformFnMap::Instance().GetNullable(
-            std::make_pair(actual_kernel_key, expected_kernel_key));
+        DataTransformFnMap::Instance().GetNullable(kernel_pair);
     if (trans_fun) {
       auto input_vars = this->InputVars();
       // TODO(qijun) filter the input vars that do not need to be transformed
@@ -437,22 +455,18 @@ void OperatorWithKernel::Run(const Scope& scope,
       }
 
       if (!need_trans.empty()) {
-        // TODO(qijun) get appropriate DeviceContext from DeviceContext pool
-        platform::DeviceContext* trans_dev_ctx = nullptr;
-        std::vector<platform::DeviceContext*> trans_dev_ctx_vec{trans_dev_ctx};
+        auto trans_dev_ctx = GetDeviceContext(kernel_pair);
 
         // Wait for transform starting
         dev_ctx->Wait();
 
         for (auto var_name : need_trans) {
-          (*trans_fun)(trans_dev_ctx_vec, *(scope.FindVar(var_name)),
+          (*trans_fun)(trans_dev_ctx, *(scope.FindVar(var_name)),
                        scope.FindVar(var_name + framework::KernelTypeToString(
                                                     expected_kernel_key)));
         }
         // Wait for data transform finishing
-        for (auto ctx : trans_dev_ctx_vec) {
-          ctx->Wait();
-        }
+        trans_dev_ctx->Wait();
       }
     }
   }
-- 
GitLab