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

CMakeLists.txt

@@ -55,6 +55,7 @@ option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
+option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)

cmake/configure.cmake

@@ -28,6 +28,10 @@ if(NOT WITH_TIMER)
     add_definitions(-DPADDLE_DISABLE_TIMER)
 endif(NOT WITH_TIMER)
 
+if(USE_EIGEN_FOR_BLAS)
+    add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
+endif(USE_EIGEN_FOR_BLAS)
+
 if(NOT WITH_PROFILER)
     add_definitions(-DPADDLE_DISABLE_PROFILER)
 endif(NOT WITH_PROFILER)

cmake/cudnn.cmake

@@ -2,7 +2,7 @@ if(NOT WITH_GPU)
     return()
 endif()
 
-set(CUDNN_ROOT "" CACHE PATH "CUDNN ROOT")
+set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT")
 find_path(CUDNN_INCLUDE_DIR cudnn.h
     PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include
     $ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/include ${CUDA_TOOLKIT_INCLUDE}

doc/api/v2/config/layer.rst

@@ -362,6 +362,11 @@ trans
 ..  autoclass:: paddle.v2.layer.trans
     :noindex:
 
+scale_shift
+-----------
+..  autoclass:: paddle.v2.layer.scale_shift
+    :noindex:
+
 Sampling Layers
 ===============
 

paddle/function/CMakeLists.txt

@@ -4,6 +4,10 @@ file(GLOB cpp_files . *Op.cpp)
 list(APPEND h_files Function.h)
 list(APPEND cpp_files Function.cpp)
 list(APPEND cpp_files BufferArg.cpp)
+list(APPEND cpp_files GemmFunctor.cpp)
+if(USE_EIGEN_FOR_BLAS)
+  list(APPEND cpp_files EigenGemm.cpp)
+endif(USE_EIGEN_FOR_BLAS)
 
 if(WITH_GPU)
     file(GLOB cu_files . *OpGpu.cu)

paddle/function/DepthwiseConvOp.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "DepthwiseConvOp.h"
 #include "ConvOp.h"
-#include "GemmFunctor.h"
 
 namespace paddle {
 

paddle/function/DepthwiseConvOpGpu.cu

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "DepthwiseConvOp.h"
-#include "GemmFunctor.h"
 #include "paddle/math/BaseMatrix.h"
 
 namespace paddle {

paddle/function/EigenGemm.cpp

+/* 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 <glog/logging.h>
+#include "unsupported/Eigen/CXX11/Tensor"
+
+namespace paddle {
+
+template <class T>
+struct EigenBlasGemm {
+  typedef Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, int>,
+                           Eigen::Aligned>
+      Matrix;
+
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+    Eigen::array<int, 2> sizeA;
+    if (transA) {
+      sizeA[0] = K;
+      sizeA[1] = M;
+      CHECK_EQ(M, lda);
+    } else {
+      sizeA[0] = M;
+      sizeA[1] = K;
+      CHECK_EQ(K, lda);
+    }
+    Eigen::array<int, 2> sizeB;
+    if (transB) {
+      sizeB[0] = N;
+      sizeB[1] = K;
+      CHECK_EQ(K, ldb);
+    } else {
+      sizeB[0] = K;
+      sizeB[1] = N;
+      CHECK_EQ(N, ldb);
+    }
+    Eigen::array<int, 2> sizeC;
+    sizeC[0] = M;
+    sizeC[1] = N;
+    CHECK_EQ(N, ldc);
+
+    const Matrix a(const_cast<T*>(A), sizeA);
+    const Matrix b(const_cast<T*>(B), sizeB);
+    Matrix c(C, sizeC);
+
+    typedef typename Eigen::Tensor<T, 2>::DimensionPair DimPair;
+    Eigen::array<DimPair, 1> dims;
+    dims[0] = DimPair(1, 0);
+    dims[0].first = transA ? 0 : 1;
+    dims[0].second = transB ? 1 : 0;
+
+    Eigen::DefaultDevice device;
+    if (alpha == T(1) && beta == T(0)) {
+      c.device(device) = a.contract(b, dims);
+    } else if (alpha == T(1) && beta == T(1)) {
+      c.device(device) += a.contract(b, dims);
+    } else {
+      c.device(device) = alpha * a.contract(b, dims) + beta * c;
+    }
+  }
+};
+
+#ifdef PADDLE_TYPE_DOUBLE
+template class EigenBlasGemm<double>;
+#else
+template class EigenBlasGemm<float>;
+#endif
+
+}  // namespace paddle

paddle/function/GemmConvOp.cpp

@@ -85,7 +85,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -108,8 +107,8 @@ public:
         int M = outputChannels / groups_;
         int N = outputHeight * outputWidth;
         int K = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasNoTrans,
+        BlasGemm<Device, real>::compute(false,
+                                        false,
                                         M,
                                         N,
                                         K,
@@ -188,8 +187,6 @@ public:
     }
 
     Col2ImFunctor<kCFO, Device, real> col2im;
-    GemmFunctor<Device, real> gemm;
-
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -205,8 +202,8 @@ public:
           colData = inputGrad + g * inputOffset;
           scale = 1.0f;
         }
-        gemm(CblasTrans,
-             CblasNoTrans,
+        BlasGemm<Device, real>::compute(true,
+                                        false,
                                         M,
                                         N,
                                         K,
@@ -299,7 +296,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -321,8 +317,8 @@ public:
         int M = outputChannels / groups_;
         int K = outputHeight * outputWidth;
         int N = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasTrans,
+        BlasGemm<Device, real>::compute(false,
+                                        true,
                                         M,
                                         N,
                                         K,

paddle/function/GemmFunctor.cpp

+/* 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 "GemmFunctor.h"
+#include "paddle/math/MathFunctions.h"
+
+namespace paddle {
+
+template <class T>
+struct BlasGemm<DEVICE_TYPE_CPU, T> {
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+#ifdef PADDLE_USE_EIGEN_FOR_BLAS
+    EigenBlasGemm<T>::compute(
+        transA, transB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
+#else
+    gemm<T>(transA == false ? CblasNoTrans : CblasTrans,
+            transB == false ? CblasNoTrans : CblasTrans,
+            M,
+            N,
+            K,
+            alpha,
+            A,
+            lda,
+            B,
+            ldb,
+            beta,
+            C,
+            ldc);
+#endif
+  }
+};
+
+template <class T>
+struct BlasGemm<DEVICE_TYPE_GPU, T> {
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+    hl_matrix_mul((T*)A,
+                  transA == false ? HPPL_OP_N : HPPL_OP_T,
+                  (T*)B,
+                  transB == false ? HPPL_OP_N : HPPL_OP_T,
+                  C,
+                  M,
+                  N,
+                  K,
+                  alpha,
+                  beta,
+                  lda,
+                  ldb,
+                  ldc);
+  }
+};
+
+template struct BlasGemm<DEVICE_TYPE_CPU, real>;
+template struct BlasGemm<DEVICE_TYPE_GPU, real>;
+
+}  // namespace paddle

paddle/function/GemmFunctor.h

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/math/MathFunctions.h"
+#include "TensorType.h"
 
 namespace paddle {
 
@@ -24,10 +24,9 @@ namespace paddle {
 // of MatMulFunction, we need to consider the reconstruction of hl_matrix_mul
 // interface.
 template <DeviceType Device, class T>
-class GemmFunctor {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
+struct BlasGemm {
+  static void compute(const bool transA,
+                      const bool transB,
                       const int M,
                       const int N,
                       const int K,
@@ -41,11 +40,15 @@ public:
                       const int ldc);
 };
 
+// TODO(hedaoyuan): Since the definition of the real type in the Paddle
+// conflicts with the Eigen library, so compile the Eigen code can not
+// include the Paddle header file. And need an EigenBlasGemm template class
+// that does not contain the DeviceType parameter.
+// I will fix this problem and merge BlasGemm and EigenBlasGemm into one.
 template <class T>
-class GemmFunctor<DEVICE_TYPE_CPU, T> {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
+struct EigenBlasGemm {
+  static void compute(const bool transA,
+                      const bool transB,
                       const int M,
                       const int N,
                       const int K,
@@ -56,41 +59,7 @@ public:
                       const int ldb,
                       const T beta,
                       T* C,
-                  const int ldc) {
-    gemm<T>(transA, TransB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
-  }
-};
-
-template <class T>
-class GemmFunctor<DEVICE_TYPE_GPU, T> {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
-                  const int M,
-                  const int N,
-                  const int K,
-                  const T alpha,
-                  const T* A,
-                  const int lda,
-                  const T* B,
-                  const int ldb,
-                  const T beta,
-                  T* C,
-                  const int ldc) {
-    hl_matrix_mul((T*)A,
-                  transA == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
-                  (T*)B,
-                  TransB == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
-                  C,
-                  M,
-                  N,
-                  K,
-                  alpha,
-                  beta,
-                  lda,
-                  ldb,
-                  ldc);
-  }
+                      const int ldc);
 };
 
 }  // namespace paddle

paddle/gserver/layers/ScaleShiftLayer.cpp

+/* 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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * A layer applies a linear transformation to each element in each row of
+ * the input matrix. For each element, the layer first re-scale it and then
+ * adds a bias to it.
+ *
+ * \f[
+ *    y = wx + b
+ * \f]
+ *
+ * Here, w is the scale and b is the bias. Both w and b are trainable scalars.
+ *
+ */
+
+class ScaleShiftLayer : public Layer {
+protected:
+  std::unique_ptr<Weight> scale_;
+  std::unique_ptr<Weight> offset_;
+
+public:
+  explicit ScaleShiftLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(scale_shift, ScaleShiftLayer);
+
+bool ScaleShiftLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(inputLayers_.size(), 1U);
+  scale_.reset(new Weight(1, 1, parameters_[0]));
+  if (biasParameter_.get() != NULL) {
+    offset_ = std::unique_ptr<Weight>(new Weight(1, 1, biasParameter_));
+  }
+  return true;
+}
+
+void ScaleShiftLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV = getInputValue(0);
+  resetOutput(inV->getHeight(), inV->getWidth());
+  MatrixPtr outV = getOutputValue();
+  real scaleValue = scale_->getW()->getElement(0, 0);
+  outV->mulScalar(*inV, scaleValue);
+  if (offset_) {
+    real offsetValue = offset_->getW()->getElement(0, 0);
+    outV->add(offsetValue);
+  }
+}
+
+void ScaleShiftLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr inG = getInputGrad(0);
+  MatrixPtr outV = getOutputValue();
+  MatrixPtr outG = getOutputGrad();
+
+  /* Calculate the parameter gradient for the current layer */
+  if (scale_->getWGrad()) {
+    MatrixPtr rowSumMtx;
+    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
+    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij}
+    rowSumMtx->sumOfProducts(
+        /* b= */ *inV, /* c= */ *outG, /* scaleSum= */ 1, /* scaleDest= */ 0.);
+    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji}
+    scale_->getWGrad()->sumCols(
+        /* b= */ *rowSumMtx, /* scaleSum= */ 1., /* scaleDest= */ 1.);
+    scale_->getParameterPtr()->incUpdate(callback);
+  }
+  if (offset_ && offset_->getWGrad()) {
+    MatrixPtr rowSumMtx;
+    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
+    rowSumMtx->sumRows(*outG, 1., 0.);
+    offset_->getWGrad()->sumCols(*rowSumMtx, 1., 1.);
+    offset_->getParameterPtr()->incUpdate(callback);
+  }
+
+  /* Calculate the input layers error */
+  if (inG) {
+    real scaleValue = scale_->getW()->getElement(0, 0);
+    inG->add(*outG, scaleValue);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -2007,6 +2007,21 @@ TEST(Layer, RowL2NormLayer) {
   }
 }
 
+TEST(Layer, ScaleShiftLayer) {
+  const size_t batchSize = 16;
+  const size_t size = 32;
+  TestConfig config;
+  config.layerConfig.set_type("scale_shift");
+  config.layerConfig.set_size(size);
+  config.biasSize = 1;
+  config.inputDefs.push_back(
+      {INPUT_DATA, "input", /* dim= */ size, /* paraSize= */ 1});
+  config.layerConfig.add_inputs();
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "scale_shift", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/gserver/tests/test_NetworkCompare.cpp

@@ -269,7 +269,8 @@ TEST(Compare, img_conv2) {
   bool useGpu = FLAGS_use_gpu;
   double eps = FLAGS_checkgrad_eps;
   FLAGS_use_gpu = true;
-  FLAGS_checkgrad_eps = 1e-2;
+  // Sometimes, this unit test will fail with 1e-2
+  FLAGS_checkgrad_eps = 4e-2;
   compareNetwork(config_file_a, config_file_b);
   FLAGS_use_gpu = useGpu;
   FLAGS_checkgrad_eps = eps;

paddle/operators/math/math_function.cc

@@ -25,8 +25,8 @@ void gemm<platform::CPUPlace, float>(const CBLAS_TRANSPOSE transA,
                                      const float alpha, const float* A,
                                      const float* B, const float beta, float* C,
                                      platform::DeviceContext* context) {
-  int lda = K;
-  int ldb = N;
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
   cblas_sgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
               beta, C, ldc);
@@ -40,8 +40,8 @@ void gemm<platform::CPUPlace, double>(const CBLAS_TRANSPOSE transA,
                                       const double* B, const double beta,
                                       double* C,
                                       platform::DeviceContext* context) {
-  int lda = K;
-  int ldb = N;
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
   cblas_dgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
               beta, C, ldc);

paddle/operators/mul_op.cc

@@ -18,6 +18,8 @@
 namespace paddle {
 namespace operators {
 
+using framework::Tensor;
+
 class MulOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -59,10 +61,23 @@ class MulOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {}
-  std::string DebugString() const override {
-    LOG(INFO) << "MulGrad";
-    return "";
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    PADDLE_ENFORCE(x_dims[0] == out_dims[0],
+                   "Out@GRAD M X N must equal to X dims 0, M ");
+    PADDLE_ENFORCE(y_dims[1] == out_dims[1],
+                   "Out@GRAD M X N must equal to Y dims 1, N ");
+
+    x_grad->Resize(x_dims);
+    y_grad->Resize(y_dims);
   }
 };
 
@@ -72,3 +87,5 @@ class MulOpGrad : public framework::OperatorWithKernel {
 namespace ops = paddle::operators;
 REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
 REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(mul_grad,
+                       ops::MulGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/mul_op.cu

@@ -17,3 +17,5 @@
 
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(mul_grad,
+                       ops::MulGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/mul_op.h

@@ -31,18 +31,34 @@ template <typename Place, typename T>
 class MulKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair = {
-        {Eigen::IndexPair<Eigen::DenseIndex>(1, 0)}};
-    auto* input0 = context.Input<Tensor>("X");
-    auto* input1 = context.Input<Tensor>("Y");
-    auto* output = context.Output<Tensor>("Out");
-    output->mutable_data<T>(context.GetPlace());
-    auto X = EigenMatrix<T>::From(*input0);
-    auto Y = EigenMatrix<T>::From(*input1);
-    auto Z = EigenMatrix<T>::From(*output);
-    auto& place = context.GetEigenDevice<Place>();
-
-    Z.device(place) = X.contract(Y, dim_pair);
+    auto* X = context.Input<Tensor>("X");
+    auto* Y = context.Input<Tensor>("Y");
+    auto* Z = context.Output<Tensor>("Out");
+    Z->mutable_data<T>(context.GetPlace());
+    auto* device_context =
+        const_cast<platform::DeviceContext*>(context.device_context_);
+    math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
+  }
+};
+
+template <typename Place, typename T>
+class MulGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* X = ctx.Input<Tensor>("X");
+    auto* Y = ctx.Input<Tensor>("Y");
+    auto* dOut = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto* dX = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dY = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    dX->mutable_data<T>(ctx.GetPlace());
+    dY->mutable_data<T>(ctx.GetPlace());
+    auto* device_context =
+        const_cast<platform::DeviceContext*>(ctx.device_context_);
+    // dX = dOut * Y'. dX: M x K, dOut : M x N, Y : K x N
+    math::matmul<Place, T>(*dOut, false, *Y, true, 1, dX, 0, device_context);
+    // dY = X' * dOut. dY: K x N, dOut : M x N, X : M x K
+    math::matmul<Place, T>(*X, true, *dOut, false, 1, dY, 0, device_context);
   }
 };
 

python/paddle/trainer/config_parser.py

@@ -2232,6 +2232,20 @@ class ClipLayer(LayerBase):
         self.config.inputs[0].clip_conf.max = max
 
 
+@config_layer('scale_shift')
+class ScaleShiftLayer(LayerBase):
+    def __init__(self, name, inputs, bias=True, **xargs):
+        super(ScaleShiftLayer, self).__init__(
+            name, 'scale_shift', 0, inputs=inputs, **xargs)
+        config_assert(
+            len(self.inputs) == 1,
+            'ScaleShiftLayer must have one and only one input.')
+        input_layer = self.get_input_layer(0)
+        self.set_layer_size(input_layer.size)
+        self.create_input_parameter(0, 1, [1, 1])
+        self.create_bias_parameter(bias, 1)
+
+
 # key: cost type
 # value: cost class
 g_cost_map = {}

python/paddle/trainer_config_helpers/layers.py

@@ -133,6 +133,7 @@ __all__ = [
     'clip_layer',
     'slice_projection',
     'kmax_sequence_score_layer',
+    'scale_shift_layer',
 ]
 
 
@@ -230,6 +231,7 @@ class LayerType(object):
     CLIP_LAYER = 'clip'
 
     KMAX_SEQ_SCORE = 'kmax_seq_score'
+    SCALE_SHIFT_LAYER = 'scale_shift'
 
     @staticmethod
     def is_layer_type(type_name):
@@ -6210,3 +6212,43 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1):
 
     return LayerOutput(
         name, LayerType.KMAX_SEQ_SCORE, parents=[input], size=input.size)
+
+
+@wrap_name_default("scale_shift")
+@wrap_param_attr_default()
+@wrap_bias_attr_default()
+def scale_shift_layer(input, name=None, param_attr=None, bias_attr=None):
+    """
+    A layer applies a linear transformation to each element in each row of 
+    the input matrix. For each element, the layer first re-scale it and then 
+    adds a bias to it.
+
+    This layer is very like the SlopeInterceptLayer, except the scale and 
+    bias are trainable.
+
+    .. math::
+
+        y = w * x + b
+
+    .. code-block:: python
+
+        scale_shift = scale_shift_layer(input=input_layer, bias_attr=False)
+
+    :param name: The Layer Name.
+    :type name: basestring
+    :param input: The input layer.
+    :type input: LayerOutput.
+    :param param_attr: The parameter attribute of scaling.
+    :type param_attr: ParameterAttribute
+    :param bias_attr: The parameter attribute of shifting.
+    :type bias_attr: ParameterAttribute
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    Layer(
+        name=name,
+        type=LayerType.SCALE_SHIFT_LAYER,
+        inputs=Input(input.name, **param_attr.attr),
+        bias=ParamAttr.to_bias(bias_attr))
+    return LayerOutput(
+        name, LayerType.SCALE_SHIFT_LAYER, parents=[input], size=input.size)

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -8,6 +8,6 @@ test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
 test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
 test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
-test_kmax_seq_socre_layer test_seq_select_layers)
+test_kmax_seq_socre_layer test_seq_select_layers test_scale_shift_layer)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_shift_layer.protostr

+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 100
+  active_type: ""
+}
+layers {
+  name: "__scale_shift_0__"
+  type: "scale_shift"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "___scale_shift_0__.w0"
+  }
+}
+layers {
+  name: "__scale_shift_1__"
+  type: "scale_shift"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "___scale_shift_1__.w0"
+  }
+  bias_parameter_name: "___scale_shift_1__.wbias"
+}
+parameters {
+  name: "___scale_shift_0__.w0"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 1.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___scale_shift_1__.w0"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 1.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___scale_shift_1__.wbias"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+input_layer_names: "data"
+output_layer_names: "__scale_shift_0__"
+output_layer_names: "__scale_shift_1__"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "__scale_shift_0__"
+  layer_names: "__scale_shift_1__"
+  input_layer_names: "data"
+  output_layer_names: "__scale_shift_0__"
+  output_layer_names: "__scale_shift_1__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py

+from paddle.trainer_config_helpers import *
+
+data = data_layer(name='data', size=100)
+
+scale = scale_shift_layer(input=data, bias_attr=False)
+
+scale_shift = scale_shift_layer(input=data)
+
+outputs(scale, scale_shift)

python/paddle/v2/framework/tests/test_mul_op.py

 import unittest
-from op_test_util import OpTestMeta
 import numpy as np
+from gradient_checker import GradientChecker, create_op
+from op_test_util import OpTestMeta
 
 
 class TestMulOp(unittest.TestCase):
@@ -15,5 +16,19 @@ class TestMulOp(unittest.TestCase):
         self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
 
 
+class MulGradOpTest(GradientChecker):
+    def test_mul(self):
+        op = create_op("mul")
+        inputs = {
+            'X': np.random.random((32, 84)).astype("float32"),
+            'Y': np.random.random((84, 100)).astype("float32")
+        }
+        # mul op will enlarge the relative error
+        self.check_grad(
+            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5)
+
+
+# TODO(dzh,qijun) : mulgrad test case need transpose feature of blas library
+
 if __name__ == '__main__':
     unittest.main()