Merge pull request #7045 from typhoonzero/adam_selectedrows

Adam selectedrows and scatter functors

paddle/operators/adagrad_op.cc

@@ -105,48 +105,18 @@ struct SparseAdagradFunctor<platform::CPUDeviceContext, T> {
                   const framework::Tensor& learning_rate, T epsilon,
                   framework::Tensor* moment, framework::Tensor* param) {
     // 1. g_m.rows = set(g.rows)
-    auto grad_rows = grad.rows();
-    std::set<int64_t> row_set(grad_rows.begin(), grad_rows.end());
-    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
-
     auto grad_width = grad.value().dims()[1];
-    std::unique_ptr<framework::SelectedRows> grad_merge{
-        new framework::SelectedRows()};
-    grad_merge->set_rows(merge_rows);
-    grad_merge->set_height(grad.height());
-    grad_merge->mutable_value()->mutable_data<T>(
-        framework::make_ddim(
-            {static_cast<int64_t>(merge_rows.size()), grad_width}),
-        context.GetPlace());
-
-    math::SetConstant<platform::CPUDeviceContext, T> constant_functor;
-    constant_functor(context, grad_merge->mutable_value(), 0.0);
-
-    auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
-    auto* grad_data = grad.value().data<T>();
-
-    for (size_t i = 0; i < grad_rows.size(); i++) {
-      size_t grad_merge_i = FindPos(merge_rows, grad_rows[i]);
-      for (int64_t j = 0; j < grad_width; j++) {
-        grad_merge_data[grad_merge_i * grad_width + j] +=
-            grad_data[i * grad_width + j];
-      }
-    }
+    math::scatter::MergeAdd<platform::CPUDeviceContext, T> merge_func;
+    auto grad_merge = merge_func(context, grad);
+    auto& merge_rows = grad_merge.rows();
+    auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
 
     // 2. m += g_m * g_m
-    std::unique_ptr<framework::SelectedRows> grad_square{
-        new framework::SelectedRows()};
-    grad_square->set_rows(grad_merge->rows());
-    grad_square->set_height(grad_merge->height());
-    grad_square->mutable_value()->mutable_data<T>(grad_merge->value().dims(),
-                                                  context.GetPlace());
-    auto gs =
-        framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
-    auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
-    gs.device(*context.eigen_device()) = gm * gm;
+    math::scatter::Mul<platform::CPUDeviceContext, T> sqare_func;
+    auto grad_square = sqare_func(context, grad_merge, grad_merge);
 
     math::SelectedRowsAddToTensor<platform::CPUDeviceContext, T> functor;
-    functor(context, *grad_square, moment);
+    functor(context, grad_square, moment);
 
     // 3. update parameter
     auto* lr = learning_rate.data<T>();

paddle/operators/adagrad_op.cu

@@ -78,62 +78,30 @@ struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
                   const framework::Tensor& learning_rate, T epsilon,
                   framework::Tensor* moment, framework::Tensor* param) {
     // 1. g_m.rows = set(g.rows)
-    auto grad_rows = grad.rows();
-    std::set<int64_t> row_set(grad_rows.begin(), grad_rows.end());
-    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
-
     auto grad_width = grad.value().dims()[1];
-    std::unique_ptr<framework::SelectedRows> grad_merge{
-        new framework::SelectedRows()};
-    grad_merge->set_rows(merge_rows);
-    grad_merge->set_height(grad.height());
-    grad_merge->mutable_value()->mutable_data<T>(
-        framework::make_ddim(
-            {static_cast<int64_t>(merge_rows.size()), grad_width}),
-        context.GetPlace());
-
-    math::SetConstant<platform::CUDADeviceContext, T> constant_functor;
-    constant_functor(context, grad_merge->mutable_value(), 0.0);
-
-    auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
-    auto* grad_data = grad.value().data<T>();
-
-    const int block_size = 256;
-    dim3 threads(block_size, 1);
-    dim3 grid1(1, grad_rows.size());
-
-    MergeGradKernel<
-        T, 256><<<grid1, threads, 0,
-                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(grad_data, grad.rows().data(),
-                                   grad_merge_data, grad_merge->rows().data(),
-                                   grad_merge->rows().size(), grad_width);
-
+    math::scatter::MergeAdd<platform::CUDADeviceContext, T> merge_func;
+    auto grad_merge = merge_func(context, grad);
+    auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
+    auto& merge_rows = grad_merge.rows();
     // 2. m += g_m * g_m
-    std::unique_ptr<framework::SelectedRows> grad_square{
-        new framework::SelectedRows()};
-    grad_square->set_rows(grad_merge->rows());
-    grad_square->set_height(grad_merge->height());
-    grad_square->mutable_value()->mutable_data<T>(grad_merge->value().dims(),
-                                                  context.GetPlace());
-    auto gs =
-        framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
-    auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
-    gs.device(*context.eigen_device()) = gm * gm;
+    math::scatter::Mul<platform::CUDADeviceContext, T> sqare_func;
+    auto grad_square = sqare_func(context, grad_merge, grad_merge);
 
     math::SelectedRowsAddToTensor<platform::CUDADeviceContext, T> functor;
-    functor(context, *grad_square, moment);
+    functor(context, grad_square, moment);
 
     // 3. update parameter
     auto* lr = learning_rate.data<T>();
     auto* param_data = param->data<T>();
     auto* moment_data = moment->data<T>();
 
+    const int block_size = 256;
+    dim3 threads(block_size, 1);
     dim3 grid2(1, merge_rows.size());
     SparseAdagradFunctorKernel<
         T, 256><<<grid2, threads, 0,
                   reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(grad_merge_data, grad_merge->rows().data(),
+                      .stream()>>>(grad_merge_data, grad_merge.rows().data(),
                                    lr, param_data, moment_data, grad_width,
                                    epsilon);
   }

paddle/operators/adam_op.h

@@ -16,11 +16,14 @@ limitations under the License. */
 #include <math.h>  // for sqrt in CPU and CUDA
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/detail/safe_ref.h"
+#include "paddle/operators/math/selected_rows_functor.h"
 #include "paddle/platform/for_range.h"
 
 namespace paddle {
 namespace operators {
 
+namespace scatter = paddle::operators::math::scatter;
+
 template <typename T>
 struct AdamFunctor {
   T beta1_;
@@ -79,6 +82,69 @@ struct AdamFunctor {
   }
 };
 
+template <typename T>
+struct SparseAdamFunctor {
+  T beta1_;
+  T beta2_;
+  T epsilon_;
+
+  const T* beta1_pow_;
+  const T* beta2_pow_;
+  const T* moment1_;
+  T* moment1_out_;
+  const T* moment2_;
+  T* moment2_out_;
+  const T* lr_;
+  const T* grad_;
+  const T* param_;
+  T* param_out_;
+
+  const int64_t* rows_;
+  int64_t row_numel_;
+
+  SparseAdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow,
+                    const T* beta2_pow, const T* mom1, T* mom1_out,
+                    const T* mom2, T* mom2_out, const T* lr, const T* grad,
+                    const T* param, T* param_out, const int64_t* rows,
+                    int64_t row_numel)
+      : beta1_(beta1),
+        beta2_(beta2),
+        epsilon_(epsilon),
+        beta1_pow_(beta1_pow),
+        beta2_pow_(beta2_pow),
+        moment1_(mom1),
+        moment1_out_(mom1_out),
+        moment2_(mom2),
+        moment2_out_(mom2_out),
+        lr_(lr),
+        grad_(grad),
+        param_(param),
+        param_out_(param_out),
+        rows_(rows),
+        row_numel_(row_numel) {}
+
+  inline HOSTDEVICE void operator()(size_t i) const {
+    T beta1_pow = *beta1_pow_;
+    T beta2_pow = *beta2_pow_;
+    for (int64_t j = 0; j < row_numel_; ++j) {
+      T g = grad_[i * row_numel_ + j];
+      T mom1 = moment1_[rows_[i] * row_numel_ + j];
+      T mom2 = moment2_[rows_[i] * row_numel_ + j];
+      T lr = *lr_;
+      T p = param_[rows_[i] * row_numel_ + j];
+
+      lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
+      mom1 = beta1_ * mom1 + (1 - beta1_) * g;
+      mom2 = beta2_ * mom2 + (1 - beta2_) * g * g;
+      p -= lr * (mom1 / (sqrt(mom2) + epsilon_));
+
+      moment1_out_[rows_[i] * row_numel_ + j] = mom1;
+      moment2_out_[rows_[i] * row_numel_ + j] = mom2;
+      param_out_[rows_[i] * row_numel_ + j] = p;
+    }  // for col id
+  }
+};
+
 template <typename DeviceContext, typename T>
 class AdamOpKernel : public framework::OpKernel<T> {
  public:
@@ -90,7 +156,8 @@ class AdamOpKernel : public framework::OpKernel<T> {
     T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
     T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
     auto& param = Ref(ctx.Input<LoDTensor>("Param"), "Must set Param");
-    auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
+    // auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
+    auto* grad_var = ctx.InputVar("Grad");
     auto& mom1 = Ref(ctx.Input<LoDTensor>("Moment1"), "Must set Moment1");
     auto& mom2 = Ref(ctx.Input<LoDTensor>("Moment2"), "Must set Moment2");
     auto& lr =
@@ -108,18 +175,48 @@ class AdamOpKernel : public framework::OpKernel<T> {
     auto& mom2_out =
         Ref(ctx.Output<LoDTensor>("Moment2Out"), "Must set Moment1Out");
 
-    AdamFunctor<T> functor(beta1, beta2, epsilon, beta1_pow.template data<T>(),
-                           beta2_pow.template data<T>(),
-                           mom1.template data<T>(),
-                           mom1_out.template mutable_data<T>(ctx.GetPlace()),
-                           mom2.template data<T>(),
-                           mom2_out.template mutable_data<T>(ctx.GetPlace()),
-                           lr.template data<T>(), grad.template data<T>(),
-                           param.template data<T>(),
-                           param_out.template mutable_data<T>(ctx.GetPlace()));
-    platform::ForRange<DeviceContext> for_range(
-        static_cast<const DeviceContext&>(ctx.device_context()), param.numel());
-    for_range(functor);
+    if (grad_var->IsType<framework::LoDTensor>()) {
+      auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
+      AdamFunctor<T> functor(
+          beta1, beta2, epsilon, beta1_pow.template data<T>(),
+          beta2_pow.template data<T>(), mom1.template data<T>(),
+          mom1_out.template mutable_data<T>(ctx.GetPlace()),
+          mom2.template data<T>(),
+          mom2_out.template mutable_data<T>(ctx.GetPlace()),
+          lr.template data<T>(), grad.template data<T>(),
+          param.template data<T>(),
+          param_out.template mutable_data<T>(ctx.GetPlace()));
+      platform::ForRange<DeviceContext> for_range(
+          static_cast<const DeviceContext&>(ctx.device_context()),
+          param.numel());
+      for_range(functor);
+    } else if (grad_var->IsType<framework::SelectedRows>()) {
+      auto& grad =
+          Ref(ctx.Input<framework::SelectedRows>("Grad"), "Must set Grad");
+      // merge duplicated rows if any.
+      scatter::MergeAdd<DeviceContext, T> merge_func;
+      auto grad_merge =
+          merge_func(ctx.template device_context<DeviceContext>(), grad);
+      auto& grad_tensor = grad_merge.value();
+      const T* grad_data = grad_tensor.template data<T>();
+      auto* rows = grad_merge.rows().data();
+      auto row_numel = grad_tensor.numel() / grad_merge.rows().size();
+
+      SparseAdamFunctor<T> functor(
+          beta1, beta2, epsilon, beta1_pow.template data<T>(),
+          beta2_pow.template data<T>(), mom1.template data<T>(),
+          mom1_out.template mutable_data<T>(ctx.GetPlace()),
+          mom2.template data<T>(),
+          mom2_out.template mutable_data<T>(ctx.GetPlace()),
+          lr.template data<T>(), grad_data, param.template data<T>(),
+          param_out.template mutable_data<T>(ctx.GetPlace()), rows, row_numel);
+      platform::ForRange<DeviceContext> for_range(
+          static_cast<const DeviceContext&>(ctx.device_context()),
+          grad_merge.rows().size());
+      for_range(functor);
+    } else {
+      PADDLE_THROW("Variable type not supported by adam_op");
+    }
   }
 };
 

paddle/operators/math/selected_rows_functor.cc

@@ -12,8 +12,10 @@ 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/operators/math/selected_rows_functor.h"
+#include <set>
+
 #include "paddle/operators/math/math_function.h"
+#include "paddle/operators/math/selected_rows_functor.h"
 
 namespace paddle {
 namespace operators {
@@ -179,6 +181,118 @@ template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, double>;
 template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, int>;
 template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, int64_t>;
 
+// This is a separated namespace for manipulate SelectedRows typed
+// data. Like merge duplicated rows, adding two SelectedRows etc.
+//
+// Another group of functors is called "scatter updates", which means
+// use SelectedRows to update a dense tensor with different Ops, like
+// add or mul.
+namespace scatter {
+
+size_t FindPos(const std::vector<int64_t>& rows, int64_t value) {
+  return std::find(rows.begin(), rows.end(), value) - rows.begin();
+}
+
+template <typename T>
+struct MergeAdd<platform::CPUDeviceContext, T> {
+  framework::SelectedRows operator()(const platform::CPUDeviceContext& context,
+                                     const framework::SelectedRows& input) {
+    framework::SelectedRows out;
+    auto input_rows = input.rows();
+    std::set<int64_t> row_set(input_rows.begin(), input_rows.end());
+    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
+
+    auto input_width = input.value().dims()[1];
+    out.set_rows(merge_rows);
+    out.set_height(input.height());
+    out.mutable_value()->mutable_data<T>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), input_width}),
+        context.GetPlace());
+
+    math::SetConstant<platform::CPUDeviceContext, T> constant_functor;
+    constant_functor(context, out.mutable_value(), 0.0);
+
+    auto* out_data = out.mutable_value()->data<T>();
+    auto* input_data = input.value().data<T>();
+
+    for (size_t i = 0; i < input_rows.size(); i++) {
+      size_t out_i = FindPos(merge_rows, input_rows[i]);
+      for (int64_t j = 0; j < input_width; j++) {
+        out_data[out_i * input_width + j] += input_data[i * input_width + j];
+      }
+    }
+    return out;
+  }
+};
+
+template struct MergeAdd<platform::CPUDeviceContext, float>;
+template struct MergeAdd<platform::CPUDeviceContext, double>;
+template struct MergeAdd<platform::CPUDeviceContext, int>;
+template struct MergeAdd<platform::CPUDeviceContext, int64_t>;
+
+template <typename T>
+struct UpdateToTensor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context,
+                  const ScatterOps& op, const framework::SelectedRows& input1,
+                  framework::Tensor* input2) {
+    auto in1_height = input1.height();
+    auto in2_dims = input2->dims();
+    PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]);
+
+    auto& in1_value = input1.value();
+    auto& in1_rows = input1.rows();
+
+    int64_t in1_row_numel = in1_value.numel() / in1_rows.size();
+    PADDLE_ENFORCE_EQ(in1_row_numel, input2->numel() / in1_height);
+
+    auto* in1_data = in1_value.data<T>();
+    auto* input2_data = input2->data<T>();
+
+    // FIXME(typhoonzero): use macro fix the below messy code.
+    switch (op) {
+      case ScatterOps::ASSIGN:
+        INLINE_FOR2(in1_rows.size(), in1_row_numel)
+        input2_data[in1_rows[i] * in1_row_numel + j] =
+            in1_data[i * in1_row_numel + j];
+        break;
+      case ScatterOps::ADD:
+        INLINE_FOR2(in1_rows.size(), in1_row_numel)
+        input2_data[in1_rows[i] * in1_row_numel + j] +=
+            in1_data[i * in1_row_numel + j];
+        break;
+      case ScatterOps::SUB:
+        INLINE_FOR2(in1_rows.size(), in1_row_numel)
+        input2_data[in1_rows[i] * in1_row_numel + j] -=
+            in1_data[i * in1_row_numel + j];
+        break;
+      case ScatterOps::SUBBY:
+        INLINE_FOR2(in1_rows.size(), in1_row_numel)
+        input2_data[in1_rows[i] * in1_row_numel + j] =
+            in1_data[i * in1_row_numel + j] -
+            input2_data[in1_rows[i] * in1_row_numel + j];
+        break;
+      case ScatterOps::MUL:
+        INLINE_FOR2(in1_rows.size(), in1_row_numel)
+        input2_data[in1_rows[i] * in1_row_numel + j] *=
+            in1_data[i * in1_row_numel + j];
+        break;
+      case ScatterOps::DIV:
+        INLINE_FOR2(in1_rows.size(), in1_row_numel)
+        input2_data[in1_rows[i] * in1_row_numel + j] /=
+            in1_data[i * in1_row_numel + j];
+        break;
+      case ScatterOps::DIVBY:
+        INLINE_FOR2(in1_rows.size(), in1_row_numel)
+        input2_data[in1_rows[i] * in1_row_numel + j] =
+            in1_data[i * in1_row_numel + j] /
+            input2_data[in1_rows[i] * in1_row_numel + j];
+        break;
+    }
+  }
+};
+
+}  // namespace scatter
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/selected_rows_functor.cu

@@ -12,6 +12,8 @@ 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 <set>
+
 #include "paddle/operators/math/math_function.h"
 #include "paddle/operators/math/selected_rows_functor.h"
 #include "paddle/platform/cuda_helper.h"
@@ -222,6 +224,157 @@ template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, float>;
 template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, double>;
 template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, int>;
 template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, int64_t>;
+
+namespace scatter {
+
+template <typename T, int block_size>
+__global__ void MergeAddKernel(const T* input, const int64_t* input_rows,
+                               T* out, const int64_t* out_rows,
+                               size_t out_rows_size, int64_t row_numel) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+  __shared__ size_t out_idx;
+
+  if (tid == 0) {
+    for (size_t i = 0; i < out_rows_size; i++) {
+      if (input_rows[ty] == out_rows[i]) {
+        out_idx = i;
+      }
+    }
+  }
+
+  __syncthreads();
+
+  input += ty * row_numel;
+  out += out_idx * row_numel;
+  for (int index = tid; index < row_numel; index += block_size) {
+    paddle::platform::CudaAtomicAdd(out + index, input[index]);
+  }
+}
+
+template <typename T>
+struct MergeAdd<platform::CUDADeviceContext, T> {
+  framework::SelectedRows operator()(const platform::CUDADeviceContext& context,
+                                     const framework::SelectedRows& input) {
+    framework::SelectedRows out;
+    auto input_rows = input.rows();
+    std::set<int64_t> row_set(input_rows.begin(), input_rows.end());
+    std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
+
+    auto input_width = input.value().dims()[1];
+
+    out.set_rows(merge_rows);
+    out.set_height(input.height());
+    out.mutable_value()->mutable_data<T>(
+        framework::make_ddim(
+            {static_cast<int64_t>(merge_rows.size()), input_width}),
+        context.GetPlace());
+
+    math::SetConstant<platform::CUDADeviceContext, T> constant_functor;
+    constant_functor(context, out.mutable_value(), 0.0);
+
+    auto* out_data = out.mutable_value()->data<T>();
+    auto* input_data = input.value().data<T>();
+
+    const int block_size = 256;
+    dim3 threads(block_size, 1);
+    dim3 grid1(1, input_rows.size());
+
+    MergeAddKernel<
+        T, 256><<<grid1, threads, 0,
+                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                      .stream()>>>(input_data, input.rows().data(), out_data,
+                                   out.rows().data(), out.rows().size(),
+                                   input_width);
+    return out;
+  }
+};
+
+template struct MergeAdd<platform::CUDADeviceContext, float>;
+template struct MergeAdd<platform::CUDADeviceContext, double>;
+template struct MergeAdd<platform::CUDADeviceContext, int>;
+template struct MergeAdd<platform::CUDADeviceContext, int64_t>;
+
+template <typename T, int block_size>
+__global__ void UpdateToTensorKernel(const T* selected_rows,
+                                     const int64_t* rows, const ScatterOps& op,
+                                     T* tensor_out, int64_t row_numel) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+
+  selected_rows += ty * row_numel;
+  tensor_out += rows[ty] * row_numel;
+  // FIXME(typhoonzero): use macro fix the below messy code.
+  switch (op) {
+    case ScatterOps::ASSIGN:
+      for (int index = tid; index < row_numel; index += block_size) {
+        tensor_out[index] = selected_rows[index];
+      }
+      break;
+    case ScatterOps::ADD:
+      for (int index = tid; index < row_numel; index += block_size) {
+        tensor_out[index] += selected_rows[index];
+      }
+      break;
+    case ScatterOps::SUB:
+      for (int index = tid; index < row_numel; index += block_size) {
+        tensor_out[index] -= selected_rows[index];
+      }
+      break;
+    case ScatterOps::SUBBY:
+      for (int index = tid; index < row_numel; index += block_size) {
+        tensor_out[index] = selected_rows[index] - tensor_out[index];
+      }
+      break;
+    case ScatterOps::MUL:
+      for (int index = tid; index < row_numel; index += block_size) {
+        tensor_out[index] *= selected_rows[index];
+      }
+      break;
+    case ScatterOps::DIV:
+      for (int index = tid; index < row_numel; index += block_size) {
+        tensor_out[index] /= selected_rows[index];
+      }
+      break;
+    case ScatterOps::DIVBY:
+      for (int index = tid; index < row_numel; index += block_size) {
+        tensor_out[index] = selected_rows[index] / tensor_out[index];
+      }
+      break;
+  }
+}
+
+template <typename T>
+struct UpdateToTensor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
+                  const ScatterOps& op, const framework::SelectedRows& input1,
+                  framework::Tensor* input2) {
+    // NOTE: Use SelectedRowsAddToTensor for better performance
+    //       no additional MergeAdd called.
+    MergeAdd<platform::CUDADeviceContext, T> merge_func;
+    auto merged_in1 = merge_func(context, input1);
+
+    auto in1_height = merged_in1.height();
+    auto in2_dims = input2->dims();
+    PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]);
+
+    auto& in1_value = merged_in1.value();
+    auto& in1_rows = merged_in1.rows();
+
+    int64_t in1_row_numel = in1_value.numel() / in1_rows.size();
+    PADDLE_ENFORCE_EQ(in1_row_numel, input2->numel() / in1_height);
+
+    auto* in1_data = in1_value.template data<T>();
+    auto* in2_data = input2->data<T>();
+
+    dim3 threads(platform::PADDLE_CUDA_NUM_THREADS, 1);
+    dim3 grid(1, in1_rows.size());
+    UpdateToTensorKernel<T, platform::PADDLE_CUDA_NUM_THREADS><<<
+        grid, threads, 0, context.stream()>>>(in1_data, in1_rows.data(), op,
+                                              in2_data, in1_row_numel);
+  }
+};
+}  // namespace scatter
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/operators/math/selected_rows_functor.h

@@ -12,9 +12,14 @@ 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/eigen.h"
 #include "paddle/framework/selected_rows.h"
 #include "paddle/platform/device_context.h"
 
+#define INLINE_FOR2(sizei, sizej)     \
+  for (int64_t i = 0; i < sizei; i++) \
+    for (int64_t j = 0; j < sizej; j++)
+
 namespace paddle {
 namespace operators {
 namespace math {
@@ -52,6 +57,78 @@ struct SelectedRowsAddToTensor {
                   framework::Tensor* input2);
 };
 
+namespace scatter {
+// functors for manuplating SelectedRows data
+template <typename DeviceContext, typename T>
+struct MergeAdd {
+  // unary functor, merge by adding duplicated rows in
+  // the input SelectedRows object.
+  framework::SelectedRows operator()(const DeviceContext& context,
+                                     const framework::SelectedRows& input);
+};
+
+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
+template <typename DeviceContext, typename T>
+struct UpdateToTensor {
+  void operator()(const DeviceContext& context, const ScatterOps& op,
+                  const framework::SelectedRows& input1,
+                  framework::Tensor* input2);
+};
+
+}  // namespace scatter
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

python/paddle/v2/fluid/tests/test_adam_op.py

@@ -1,6 +1,8 @@
 import unittest
 import numpy as np
 from op_test import OpTest
+from paddle.v2.fluid import core
+from paddle.v2.fluid.op import Operator
 
 
 class TestAdamOp1(OpTest):
@@ -176,5 +178,124 @@ def adam_step(inputs, attributes):
     return param_out, moment1_out, moment2_out
 
 
+def adam_step_sparse(inputs, attributes, height, rows, row_numel, np_grad):
+    '''
+    Simulate one step of the adam optimizer
+    :param inputs: dict of inputs
+    :param attributes: dict of attributes
+    :return tuple: tuple of output param, moment1, moment2,
+    beta1 power accumulator and beta2 power accumulator
+    '''
+    param = inputs['Param']
+    # grad = inputs['Grad']
+    moment1 = inputs['Moment1']
+    moment2 = inputs['Moment2']
+    lr = inputs['LearningRate']
+    beta1_pow = inputs['Beta1Pow']
+    beta2_pow = inputs['Beta2Pow']
+
+    beta1 = attributes['beta1']
+    beta2 = attributes['beta2']
+    epsilon = attributes['epsilon']
+
+    moment1_out = np.zeros(shape=[height, row_numel])
+    moment2_out = np.zeros(shape=[height, row_numel])
+    param_out = np.zeros(shape=[height, row_numel])
+
+    for idx, row_id in enumerate(rows):
+        moment1_out[row_id] = beta1 * moment1[row_id] + (1 - beta1
+                                                         ) * np_grad[idx]
+        moment2_out[row_id] = beta2 * moment2[row_id] + (
+            1 - beta2) * np.square(np_grad[idx])
+        lr_t = lr * np.sqrt(1 - beta2_pow) / (1 - beta1_pow)
+        param_out[row_id] = param[row_id] - lr_t * (moment1_out[row_id] / (
+            np.sqrt(moment2_out[row_id]) + epsilon))
+    return param_out, moment1_out, moment2_out
+
+
+class TestSparseAdamOp(unittest.TestCase):
+    def setup(self, scope, place):
+        beta1 = 0.78
+        beta2 = 0.836
+        epsilon = 1e-4
+
+        height = 10
+        rows = [0, 4, 7]
+        self.rows = rows
+        row_numel = 12
+        self.row_numel = row_numel
+        self.dense_inputs = {
+            "Param": np.full((height, row_numel), 5.0).astype("float32"),
+            "Moment1": np.full((height, row_numel), 5.0).astype("float32"),
+            "Moment2": np.full((height, row_numel), 5.0).astype("float32"),
+            'Beta1Pow': np.array([beta1**10]).astype("float32"),
+            'Beta2Pow': np.array([beta2**10]).astype("float32"),
+            "LearningRate": np.full((1), 2.0).astype("float32")
+        }
+        self.attrs = {'epsilon': epsilon, 'beta1': beta1, 'beta2': beta2}
+
+        grad_selected_rows = scope.var('Grad').get_selected_rows()
+        grad_selected_rows.set_height(height)
+        grad_selected_rows.set_rows(rows)
+        np_array = np.ones((len(rows), row_numel)).astype("float32")
+        np_array[0, 0] = 2.0
+        np_array[2, 8] = 4.0
+
+        grad_tensor = grad_selected_rows.get_tensor()
+        grad_tensor.set(np_array, place)
+
+        self.sparse_inputs = ["Grad"]
+
+        param_out, mom1, mom2 = adam_step_sparse(
+            self.dense_inputs, self.attrs, height, rows, row_numel, np_array)
+        self.outputs = {
+            "ParamOut": param_out,
+            "Moment1Out": mom1,
+            "Moment2Out": mom2
+        }
+
+    def check_with_place(self, place):
+        scope = core.Scope()
+        self.setup(scope, place)
+
+        op_args = dict()
+        for key, np_array in self.dense_inputs.iteritems():
+            var = scope.var(key).get_tensor()
+            var.set(np_array, place)
+            op_args[key] = key
+        for s in self.sparse_inputs:
+            op_args[s] = s
+        for s in self.outputs:
+            var = scope.var(s).get_tensor()
+            var.set(self.outputs[s], place)
+            op_args[s] = s
+        for k in self.attrs:
+            op_args[k] = self.attrs[k]
+
+        # create and run sgd operator
+        adam_op = Operator("adam", **op_args)
+        adam_op.run(scope, place)
+
+        for key, np_array in self.outputs.iteritems():
+            out_var = scope.var(key).get_tensor()
+            actual = np.array(out_var)
+            actual = actual.reshape([actual.size])
+            np_array = np_array.reshape([np_array.size])
+            for idx, row_id in enumerate(self.rows):
+                j = 0
+                while j < self.row_numel:
+                    pos = row_id * self.row_numel + j
+                    self.assertLess((actual[pos] - np_array[pos]) / actual[pos],
+                                    0.00001)
+                    j += 1
+
+    def test_sparse_sgd(self):
+        places = [core.CPUPlace()]
+        if core.is_compile_gpu():
+            places.append(core.CUDAPlace(0))
+        for place in places:
+            self.check_with_place(place)
+
+
 if __name__ == "__main__":
     unittest.main()