Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into...

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into dist-table-support-optimizer-regular

paddle/fluid/operators/adagrad_op.cc

@@ -119,8 +119,8 @@ struct SparseAdagradFunctor<platform::CPUDeviceContext, T> {
     auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
 
     // 2. m += g_m * g_m
-    math::scatter::Mul<platform::CPUDeviceContext, T> sqare_func;
-    auto grad_square = sqare_func(context, grad_merge, grad_merge);
+    auto grad_square =
+        SquareSelectedRows<platform::CPUDeviceContext, T>(context, grad_merge);
 
     math::SelectedRowsAddToTensor<platform::CPUDeviceContext, T> functor;
     functor(context, grad_square, moment);

paddle/fluid/operators/adagrad_op.cu

@@ -84,8 +84,8 @@ struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
     auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
     framework::Vector<int64_t> merge_rows(grad_merge.rows());
     // 2. m += g_m * g_m
-    math::scatter::Mul<platform::CUDADeviceContext, T> sqare_func;
-    auto grad_square = sqare_func(context, grad_merge, grad_merge);
+    auto grad_square =
+        SquareSelectedRows<platform::CUDADeviceContext, T>(context, grad_merge);
 
     math::SelectedRowsAddToTensor<platform::CUDADeviceContext, T> functor;
     functor(context, grad_square, moment);

paddle/fluid/operators/adagrad_op.h

@@ -28,6 +28,20 @@ struct SparseAdagradFunctor {
                   framework::Tensor *moment, framework::Tensor *param);
 };
 
+template <typename DeviceContext, typename T>
+framework::SelectedRows SquareSelectedRows(
+    const DeviceContext &context, const framework::SelectedRows &input) {
+  framework::SelectedRows out;
+  out.set_rows(input.rows());
+  out.set_height(input.height());
+  out.mutable_value()->mutable_data<T>(input.value().dims(),
+                                       context.GetPlace());
+  auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
+  auto e_in = framework::EigenVector<T>::Flatten(input.value());
+  e_out.device(*context.eigen_device()) = e_in.square();
+  return out;
+}
+
 template <typename DeviceContext, typename T>
 class AdagradOpKernel : public framework::OpKernel<T> {
  public:

paddle/fluid/operators/math/jit_kernel_blas.cc

@@ -168,9 +168,11 @@ class VAddKernelImpl : public VAddKernel<T> {
 #endif
     this->Compute = VAddRefer<T>;
   }
+#ifdef PADDLE_WITH_XBYAK
 
  private:
   std::unique_ptr<gen::VVVJitCode> jitcode_{nullptr};
+#endif
 };
 
 #ifdef PADDLE_WITH_XBYAK
@@ -210,9 +212,11 @@ class VAddReluKernelImpl : public VAddReluKernel<T> {
 #endif
     this->Compute = VAddReluRefer<T>;
   }
+#ifdef PADDLE_WITH_XBYAK
 
  private:
   std::unique_ptr<gen::VVVJitCode> jitcode_{nullptr};
+#endif
 };
 
 #ifdef PADDLE_WITH_XBYAK

paddle/fluid/operators/math/selected_rows_functor.h

@@ -88,57 +88,6 @@ struct MergeAdd {
                   framework::SelectedRows* output);
 };
 
-template <typename DeviceContext, typename T>
-struct Add {
-  framework::SelectedRows operator()(const DeviceContext& context,
-                                     const framework::SelectedRows& input1,
-                                     const framework::SelectedRows& input2) {
-    framework::SelectedRows out;
-    out.set_rows(input1.rows());
-    out.set_height(input1.height());
-    out.mutable_value()->mutable_data<T>(input1.value().dims(),
-                                         context.GetPlace());
-    auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
-    auto e_in1 = framework::EigenVector<T>::Flatten(input1.value());
-    auto e_in2 = framework::EigenVector<T>::Flatten(input2.value());
-    e_out.device(*context.eigen_device()) = e_in1 + e_in2;
-    return out;
-  }
-};
-
-template <typename DeviceContext, typename T>
-struct Mul {
-  // multiply two SelectedRows
-  framework::SelectedRows operator()(const DeviceContext& context,
-                                     const framework::SelectedRows& input1,
-                                     const framework::SelectedRows& input2) {
-    framework::SelectedRows out;
-    out.set_rows(input1.rows());
-    out.set_height(input1.height());
-    out.mutable_value()->mutable_data<T>(input1.value().dims(),
-                                         context.GetPlace());
-    auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
-    auto e_in1 = framework::EigenVector<T>::Flatten(input1.value());
-    auto e_in2 = framework::EigenVector<T>::Flatten(input2.value());
-    e_out.device(*context.eigen_device()) = e_in1 * e_in2;
-    return out;
-  }
-  // multiply scalar to SelectedRows
-  framework::SelectedRows operator()(const DeviceContext& context,
-                                     const framework::SelectedRows& input1,
-                                     const T input2) {
-    framework::SelectedRows out;
-    out.set_rows(input1.rows());
-    out.set_height(input1.height());
-    out.mutable_value()->mutable_data<T>(input1.value().dims(),
-                                         context.GetPlace());
-    auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
-    auto e_in1 = framework::EigenVector<T>::Flatten(input1.value());
-    e_out.device(*context.eigen_device()) = input2 * e_in1;
-    return out;
-  }
-};
-
 enum class ScatterOps { ASSIGN, ADD, SUB, SUBBY, MUL, DIV, DIVBY };
 
 // out = seleted_rows_in / tensor

python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py

@@ -16,6 +16,7 @@ from __future__ import print_function
 
 import paddle.dataset.conll05 as conll05
 import paddle.fluid as fluid
+import paddle.fluid.core as core
 import unittest
 import paddle
 import numpy as np
@@ -177,6 +178,7 @@ class TestCRFModel(unittest.TestCase):
     def test_update_sparse_parameter_all_reduce(self):
         build_strategy = fluid.BuildStrategy()
         build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
+        if core.is_compiled_with_cuda():
             self.check_network_convergence(
                 is_sparse=True, build_strategy=build_strategy, use_cuda=True)
         self.check_network_convergence(
@@ -185,6 +187,7 @@ class TestCRFModel(unittest.TestCase):
     def test_update_dense_parameter_all_reduce(self):
         build_strategy = fluid.BuildStrategy()
         build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
+        if core.is_compiled_with_cuda():
             self.check_network_convergence(
                 is_sparse=False, build_strategy=build_strategy, use_cuda=True)
         self.check_network_convergence(
@@ -193,6 +196,7 @@ class TestCRFModel(unittest.TestCase):
     def test_update_sparse_parameter_reduce(self):
         build_strategy = fluid.BuildStrategy()
         build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
+        if core.is_compiled_with_cuda():
             self.check_network_convergence(
                 is_sparse=True, build_strategy=build_strategy, use_cuda=True)
         self.check_network_convergence(
@@ -201,6 +205,7 @@ class TestCRFModel(unittest.TestCase):
     def test_update_dense_parameter_reduce(self):
         build_strategy = fluid.BuildStrategy()
         build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
+        if core.is_compiled_with_cuda():
             self.check_network_convergence(
                 is_sparse=False, build_strategy=build_strategy, use_cuda=True)
         self.check_network_convergence(