Merge branch 'develop' of github.com:baidu/Paddle into feature/parallel_for_unittest

Author: reyoung

paddle/framework/CMakeLists.txt

@@ -32,10 +32,12 @@ cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool)
 cc_library(scope SRCS scope.cc DEPS glog threadpool)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
-cc_library(device_data_transform SRCS device_data_transform.cc DEPS tensor)
+cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor)
+cc_library(data_type_transform SRCS data_type_transform.cc DEPS tensor)
+cc_library(data_layout_transform SRCS data_layout_transform.cc DEPS tensor math_function)
 
-cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor framework_proto selected_rows device_data_transform)
-cc_test(data_transform_test SRCS data_transform_test.cc DEPS data_transform device_context)
+cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor
+        framework_proto selected_rows data_device_transform data_type_transform data_layout_transform)
 
 cc_library(attribute SRCS attribute.cc DEPS framework_proto)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc
@@ -80,5 +82,5 @@ cc_test(init_test SRCS init_test.cc DEPS init)
 
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
-nv_test(device_data_transform_test SRCS device_data_transform_test.cu
+nv_test(data_device_transform_test SRCS data_device_transform_test.cu
         DEPS operator op_registry init math_function)

paddle/framework/device_data_transform.cc renamed to paddle/framework/data_device_transform.cc

@@ -11,7 +11,7 @@ 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/framework/device_data_transform.h"
+#include "paddle/framework/data_device_transform.h"
 
 namespace paddle {
 namespace framework {

paddle/framework/device_data_transform.h renamed to paddle/framework/data_device_transform.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "paddle/platform/enforce.h"
 
 #include <iostream>
 #include "paddle/platform/enforce.h"

paddle/framework/device_data_transform_test.cu renamed to paddle/framework/data_device_transform_test.cu

@@ -0,0 +1,82 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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/framework/data_layout_transform.h"
+
+#include "paddle/framework/tensor.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace framework {
+
+struct CastDataLayout {
+  CastDataLayout(const platform::DeviceContext* ctx,
+                 const std::vector<int>& axis, const framework::Tensor& in,
+                 framework::Tensor* out)
+      : in_(in), out_(out), ctx_(ctx), axis_(axis) {}
+  const framework::Tensor in_;
+  framework::Tensor* out_;
+  const platform::DeviceContext* ctx_;
+  const std::vector<int> axis_;
+
+  template <typename T>
+  void operator()() {
+    auto place = ctx_->GetPlace();
+
+    if (platform::is_cpu_place(place)) {
+      operators::math::Transpose<platform::CPUDeviceContext, T, 4> trans4;
+      auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
+      trans4(*context, in_, out_, axis_);
+    } else {
+      PADDLE_THROW("Unsupport CPU <-> GPU!");
+    }
+  }
+};
+
+void TransDataLayout(const std::vector<int>& axis,
+                     const platform::DeviceContext* ctx,
+                     const KernelTypePair& kernel_pair, const Variable& in,
+                     Variable* out) {
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only support Tensor transform!.");
+  PADDLE_ENFORCE(
+      platform::places_are_same_class(kernel_pair.first.place_,
+                                      kernel_pair.second.place_),
+      "TransDataLayout only support DataLayout transform on same place!");
+  PADDLE_ENFORCE(kernel_pair.first.data_type_ == kernel_pair.second.data_type_,
+                 "TransDataLayout only support Datatype are same!");
+
+  auto src = in.Get<Tensor>();
+  auto* dst = out->GetMutable<Tensor>();
+  PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!");
+
+  auto src_dim = src.dims();
+  std::vector<int64_t> dst_dim;
+
+  dst_dim.resize(axis.size());
+  for (size_t i = 0; i < axis.size(); i++) {
+    dst_dim[i] = src_dim[axis[i]];
+  }
+
+  dst->Resize(make_ddim(dst_dim));
+  auto place = kernel_pair.second.place_;
+  dst->mutable_data(place, src.type());
+
+  auto src_type = kernel_pair.first.data_type_;
+  framework::VisitDataType(src_type, CastDataLayout(ctx, axis, src, dst));
+
+  dst->set_layout(kernel_pair.second.data_layout_);
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_layout.h

@@ -0,0 +1,31 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 "paddle/framework/op_kernel_type.h"
+#include "paddle/framework/variable.h"
+
+namespace paddle {
+namespace framework {
+
+using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
+
+void TransDataLayout(const std::vector<int>& axis,
+                     const platform::DeviceContext* ctx,
+                     const KernelTypePair& kernel_pair, const Variable& in,
+                     Variable* out);
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_layout_transform.cc

@@ -11,22 +11,14 @@ 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 <functional>
 
 #include "paddle/framework/data_transform.h"
-#include "paddle/framework/device_data_transform.h"
-#include "paddle/framework/lod_tensor.h"
-#include "paddle/framework/selected_rows.h"
-#include "paddle/platform/device_context.h"
+
+#include "paddle/framework/data_device_transform.h"
 
 namespace paddle {
 namespace framework {
 
-DataTransformFnMap& DataTransformFnMap::Instance() {
-  static DataTransformFnMap data_transform_map;
-  return data_transform_map;
-}
-
 Tensor* DataTransform(const OpKernelType& expected_kernel_type,
                       const OpKernelType& kernel_type_for_var,
                       const Tensor& input_tensor) {
@@ -58,134 +50,5 @@ void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
   }
 }
 
-auto KernelFP32 = OpKernelType(proto::DataType::FP32, platform::CPUPlace(),
-                               DataLayout::kNHWC, LibraryType::kPlain);
-
-auto KernelFP64 = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
-                               DataLayout::kNHWC, LibraryType::kPlain);
-
-auto KernelNHWC = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
-                               DataLayout::kNHWC, LibraryType::kPlain);
-
-auto KernelNCHW = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
-                               DataLayout::kNCHW, LibraryType::kPlain);
-
-// TODO(dzhwinter): Only for testing multiple op kernel.
-// Dummy transform function for library_type
-// should be removed.
-auto KernelPlain = OpKernelType(proto::DataType::FP32, platform::CUDAPlace(0),
-                                DataLayout::kAnyLayout, LibraryType::kPlain);
-
-auto KernelCUDNN = OpKernelType(proto::DataType::FP32, platform::CUDAPlace(0),
-                                DataLayout::kAnyLayout, LibraryType::kCUDNN);
-
-void DummyTrans(const platform::DeviceContext* ctx,
-                const KernelTypePair& kernel_pair, const Variable& in,
-                Variable* out) {
-  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
-  PADDLE_ENFORCE(
-      platform::places_are_same_class(kernel_pair.first.place_,
-                                      kernel_pair.second.place_),
-      "TransDataType Only Support DataType transform on same place!");
-  auto src = in.Get<Tensor>();
-  auto* dst = out->GetMutable<Tensor>();
-  *dst = src;
-}
-
-void TransDataType(const platform::DeviceContext* ctx,
-                   const KernelTypePair& kernel_pair, const Variable& in,
-                   Variable* out) {
-  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
-  PADDLE_ENFORCE(
-      platform::places_are_same_class(kernel_pair.first.place_,
-                                      kernel_pair.second.place_),
-      "TransDataType Only Support DataType transform on same place!");
-
-  auto src = in.Get<Tensor>();
-  auto* dst = out->GetMutable<Tensor>();
-
-  auto dims = src.dims();
-  dst->Resize(dims);
-  auto dst_type = kernel_pair.second.data_type_;
-  auto src_type = kernel_pair.first.data_type_;
-
-  switch (src_type) {
-    case proto::DataType::FP32:
-      framework::VisitDataType(dst_type, CastDataType<float>(src, dst, ctx));
-      break;
-    case proto::DataType::FP64:
-      framework::VisitDataType(dst_type, CastDataType<double>(src, dst, ctx));
-      break;
-    case proto::DataType::INT32:
-      framework::VisitDataType(dst_type, CastDataType<int>(src, dst, ctx));
-      break;
-    case proto::DataType::INT64:
-      framework::VisitDataType(dst_type, CastDataType<int64_t>(src, dst, ctx));
-      break;
-    case proto::DataType::BOOL:
-      framework::VisitDataType(dst_type, CastDataType<bool>(src, dst, ctx));
-      break;
-    default:
-      PADDLE_THROW("Not support type %d", src_type);
-  }
-}
-
-void TransDataLayout(const std::vector<int>& axis,
-                     const platform::DeviceContext* ctx,
-                     const KernelTypePair& kernel_pair, const Variable& in,
-                     Variable* out) {
-  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only support Tensor transform!.");
-  PADDLE_ENFORCE(
-      platform::places_are_same_class(kernel_pair.first.place_,
-                                      kernel_pair.second.place_),
-      "TransDataLayout only support DataLayout transform on same place!");
-  PADDLE_ENFORCE(kernel_pair.first.data_type_ == kernel_pair.second.data_type_,
-                 "TransDataLayout only support Datatype are same!");
-
-  auto src = in.Get<Tensor>();
-  auto* dst = out->GetMutable<Tensor>();
-  PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!");
-
-  auto src_dim = src.dims();
-  std::vector<int64_t> dst_dim;
-
-  dst_dim.resize(axis.size());
-  for (size_t i = 0; i < axis.size(); i++) {
-    dst_dim[i] = src_dim[axis[i]];
-  }
-
-  dst->Resize(make_ddim(dst_dim));
-  auto place = kernel_pair.second.place_;
-  dst->mutable_data(place, src.type());
-
-  auto src_type = kernel_pair.first.data_type_;
-  framework::VisitDataType(src_type, CastDataLayout(ctx, axis, src, dst));
-
-  dst->set_layout(kernel_pair.second.data_layout_);
-}
-
 }  // namespace framework
 }  // namespace paddle
-
-namespace f = paddle::framework;
-
-namespace {
-std::vector<int> NHWC2NCHW = {0, 3, 1, 2};
-std::vector<int> NCHW2NHWC = {0, 2, 3, 1};
-}
-
-REGISTER_DATA_TRANSFORM_FN(f::KernelFP32, f::KernelFP64, f::TransDataType);
-REGISTER_DATA_TRANSFORM_FN(f::KernelPlain, f::KernelCUDNN, f::DummyTrans);
-REGISTER_DATA_TRANSFORM_FN(f::KernelCUDNN, f::KernelPlain, f::DummyTrans);
-REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW,
-                           std::bind(f::TransDataLayout, NHWC2NCHW,
-                                     std::placeholders::_1,
-                                     std::placeholders::_2,
-                                     std::placeholders::_3,
-                                     std::placeholders::_4));
-REGISTER_DATA_TRANSFORM_FN(f::KernelNCHW, f::KernelNHWC,
-                           std::bind(f::TransDataLayout, NCHW2NHWC,
-                                     std::placeholders::_1,
-                                     std::placeholders::_2,
-                                     std::placeholders::_3,
-                                     std::placeholders::_4));