Merge pull request #12878 from tensor-tang/feature/op/attention_lstm

Add attention lstm cpu forward

CMakeLists.txt

@@ -138,12 +138,6 @@ else()
     set(THIRD_PARTY_BUILD_TYPE Release)
 endif()
 
-if(WITH_MKL)
-  option(MKL_SPLIT_GEMM "PaddlePaddle MKL gemm would split to small ones" OFF)
-  if (MKL_SPLIT_GEMM)
-    add_definitions(-DPADDLE_MKL_SPLIT_GEMM)
-  endif()
-endif()
 set(WITH_MKLML ${WITH_MKL})
 if (NOT DEFINED WITH_MKLDNN)
     if (WITH_MKL AND AVX2_FOUND)

paddle/fluid/operators/attention_lstm_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+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/fluid/operators/attention_lstm_op.h"
+#include <sys/time.h>
+#include <string>
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/cpu_vec.h"
+#include "paddle/fluid/operators/math/fc_compute.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+
+void AttentionLSTMOp::InferShape(framework::InferShapeContext* ctx) const {
+  PADDLE_ENFORCE(ctx->HasInput("X"),
+                 "Input(X) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("C0"),
+                 "Input(C0) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("LSTMWeight"),
+                 "Input(LSTMWeight) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("LSTMBias"),
+                 "Input(LSTMBias) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("AttentionWeight"),
+                 "Input(AttentionWeight) of AttentionLSTM should not be null.");
+
+  PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
+                 "Output(Hidden) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("Cell"),
+                 "Output(Cell) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("AttentionedX"),
+                 "Output(AttentionedX) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("AttentionFCOut"),
+                 "Output(AttentionFCOut) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("LSTMX"),
+                 "Output(LSTMX) of AttentionLSTM should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("LSTMOUT"),
+                 "Output(LSTMOUT) of AttentionLSTM should not be null.");
+
+  auto x_dims = ctx->GetInputDim("X");
+  const int M = x_dims[1];
+  PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2.");
+
+  auto w_dims = ctx->GetInputDim("LSTMWeight");
+  const int D = w_dims[1] / 4;
+  PADDLE_ENFORCE_EQ(w_dims.size(), 2, "Input(LSTMWeight)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(w_dims[0], D + M,
+                    "LSTMWeight dims should be (%d + %d) * %d.", D + M, 4 * D);
+
+  auto b_dims = ctx->GetInputDim("LSTMBias");
+  PADDLE_ENFORCE_EQ(b_dims.size(), 2, "Input(LSTMBias)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(b_dims[0], 1, "LSTMBias dims should be 1 x %d.", 4 * D);
+  PADDLE_ENFORCE_EQ(b_dims[1], 4 * D, "LSTMBias dims should be 1 x %d.", 4 * D);
+
+  auto c_dims = ctx->GetInputDim("C0");
+  PADDLE_ENFORCE_EQ(c_dims.size(), 2, "Input(C0)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(c_dims[1], D, "C0 dims should be N x %d.", D);
+  if (ctx->HasInput("H0")) {
+    auto h_dims = ctx->GetInputDim("H0");
+    PADDLE_ENFORCE(h_dims == c_dims,
+                   "The dimension of Input(H0) and Input(C0) "
+                   "should be the same.");
+  }
+
+  auto atten_w_dims = ctx->GetInputDim("AttentionWeight");
+  PADDLE_ENFORCE_EQ(atten_w_dims.size(), 2,
+                    "Input(AttentionWeight)'s rank must be 2.");
+  PADDLE_ENFORCE_EQ(atten_w_dims[0], M + D,
+                    "AttentionWeight shapes must be (%d + %d) * 1.", M, D);
+  PADDLE_ENFORCE_EQ(atten_w_dims[1], 1,
+                    "AttentionWeight shapes must be (%d + %d) * 1.", M, D);
+  if (ctx->HasInput("AttentionBias")) {
+    auto atten_b_dims = ctx->GetInputDim("AttentionBias");
+    PADDLE_ENFORCE_EQ(atten_b_dims.size(), 2,
+                      "Input(AttentionBias)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(atten_b_dims[0], 1,
+                      "AttentionBias shapes must be 1 * 1.");
+    PADDLE_ENFORCE_EQ(atten_b_dims[1], 1,
+                      "AttentionBias shapes must be 1 * 1.");
+  }
+
+  if (ctx->HasInput("AttentionScalar")) {
+    auto dims = ctx->GetInputDim("AttentionScalar");
+    PADDLE_ENFORCE_EQ(dims.size(), 2,
+                      "Input(AttentionScalar)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(dims[0], 1, "AttentionScalar shapes must be 1 * 1.");
+    PADDLE_ENFORCE_EQ(dims[1], 1, "AttentionScalar shapes must be 1 * 1.");
+  }
+
+  if (ctx->HasInput("AttentionScalarBias")) {
+    auto dims = ctx->GetInputDim("AttentionScalarBias");
+    PADDLE_ENFORCE(
+        ctx->HasInput("AttentionScalar"),
+        "AttentionScalar should not be null when have AttentionScalarBias.");
+    PADDLE_ENFORCE_EQ(dims.size(), 2,
+                      "Input(AttentionScalarBias)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(dims[0], 1, "AttentionScalarBias shapes must be 1 * 1.");
+    PADDLE_ENFORCE_EQ(dims[1], 1, "AttentionScalarBias shapes must be 1 * 1.");
+  }
+
+  framework::DDim out_dims({x_dims[0], D});
+  ctx->SetOutputDim("Hidden", out_dims);
+  ctx->SetOutputDim("Cell", out_dims);
+  ctx->SetOutputDim("AttentionedX", {x_dims[0], 1});
+  ctx->SetOutputDim("LSTMX", {1, M});
+  ctx->SetOutputDim("LSTMOUT", {1, 4 * D});
+  // AttentionFCOut should be reshape as (maxseqlen,1) in runtime
+  ctx->ShareLoD("X", "Hidden");
+  ctx->ShareLoD("X", "Cell");
+}
+
+framework::OpKernelType AttentionLSTMOp::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
+      ctx.device_context());
+}
+
+void AttentionLSTMOpMaker::Make() {
+  AddInput("X",
+           "(LoDTensor) the input is a LodTensor, which support "
+           "variable-time length input sequence. The underlying tensor in "
+           "this LoDTensor is a matrix with shape (T X M), where T is the "
+           "total time steps in this mini-batch, M is the dim size of x.");
+  AddInput("C0",
+           "(Tensor) LSTM C0"
+           "This is a tensor with shape (N x D), where N is the batch size, D "
+           "is the gate size."
+           "C0 is necessary because of attention.");
+  AddInput("H0",
+           "(Tensor, optional) LSTM H0"
+           "This is a tensor with shape (N x D), where N is the "
+           "batch size and D is the gate size.")
+      .AsDispensable();
+  AddInput("AttentionWeight",
+           "(Tensor) the weights of attention fc. Always relu the fc result."
+           "The shape is ((M+D) x 1), where M is the dim size of x, D is the "
+           "gate size of LSTM.");
+  AddInput("AttentionBias",
+           "(Tensor, optional) the bias of attention fc."
+           "The shape is (1 x 1)")
+      .AsDispensable();
+  AddInput("AttentionScalar",
+           "(Tensor, optional) the scalar on the result of attentioned fc. "
+           "Always relu the Scalar."
+           "The shape is (1 x 1)")
+      .AsDispensable();
+  AddInput("AttentionScalarBias",
+           "(Tensor, optional) the scalar bias of attention fc."
+           "The shape is (1 x 1)")
+      .AsDispensable();
+  AddInput("LSTMWeight",
+           "(Tensor) the combined weight of LSTM"
+           " - The shape is ((D+M) x 4D), where D is the hidden gate size, M "
+           "is the dim size of x"
+           " - Weight = {W_forget, W_input, W_output, W_cell}");
+  AddInput("LSTMBias",
+           "(Tensor) the combined bias of LSTM, shape (1x4D)."
+           "Note: we should add the bias of hidden and context accorindg to "
+           "the same gate: "
+           "{B_forget, B_input, B_output, B_cell}");
+  AddOutput("Hidden",
+            "(LoDTensor) (same as LSTMOp) the hidden state of LSTM operator. "
+            "The shape is (T x D), and lod is the same with the `Input`.");
+  AddOutput("Cell",
+            "(LoDTensor) (same as LSTMOp) the cell state of LSTM operator. "
+            "The shape is (T x D), and lod is the same with the `Input`.");
+  AddOutput("AttentionedX",
+            "(Tensor) shape is (T x 1), the result after X * AttentionWeight,"
+            " where T is the total time steps in this mini-batch,"
+            " D is the hidden size.")
+      .AsIntermediate();
+  AddOutput("AttentionFCOut",
+            "(Tensor) (max_seq_len, 1), compute at each step.")
+      .AsIntermediate();
+  AddOutput("LSTMX",
+            "(Tensor) the input X of LSTM for each step."
+            "Shape is (1 x M), where M is the x frame size")
+      .AsIntermediate();
+  AddOutput(
+      "LSTMOUT",
+      "(Tensor) the output of LSTM X(1*(D+M))* weight((D+M)*4D) for each step."
+      "Shape is (1 x 4D), where M is the x frame size")
+      .AsIntermediate();
+  AddAttr<std::string>("gate_activation",
+                       "(string, default: sigmoid)"
+                       "The activation for input gate, forget gate and output "
+                       "gate, `sigmoid` by default.")
+      .SetDefault("sigmoid")
+      .InEnum({"sigmoid", "tanh", "relu", "identity"});
+  AddAttr<std::string>("cell_activation",
+                       "(string, default: tanh)"
+                       "The activation for cell output, `tanh` by defalut.")
+      .SetDefault("tanh")
+      .InEnum({"sigmoid", "tanh", "relu", "identity"});
+  AddAttr<std::string>("candidate_activation",
+                       "(string, default: tanh)"
+                       "The activation for candidate hidden state, "
+                       "`tanh` by default.")
+      .SetDefault("tanh")
+      .InEnum({"sigmoid", "tanh", "relu", "identity"});
+  AddComment(R"DOC(
+Attention Long-Short Term Memory (LSTM) Operator.
+
+Attention part:
+concat( x(seqlen * M), expand( cell_t-1(1,D) ) ) => tmp(seqlen*(M+D))
+
+tmp(seqlen*(M+D)) * fc((M+D)*1) => fcout(seqlen*1) with bias, relu
+
+fcout(seqlen*1) * scalar => fcout(seqlen*1) with bias, relu
+
+dotmul and sum pool ( fcout(seqlen*1), x(seqlen * M) ) => lstm_x_t(1, M) 
+
+LSTM part:
+use lstm_x_t as input and compute as standard LSTM.
+
+)DOC");
+}
+
+// y[i] = (x[i] + bias[0]) > 0 ? (x[i] + bias[0]) : 0;
+template <typename T>
+inline void bias_relu(const int n, const T* x, const T* bias, T* y) {
+  if (bias) {
+    for (int i = 0; i < n; ++i) {
+      y[i] = x[i] + bias[0];
+    }
+    math::vec_relu<T>(n, y, y);
+  } else {
+    math::vec_relu<T>(n, x, y);
+  }
+}
+
+template <typename DeviceContext, typename T>
+inline void vec_softmax(const math::BlasT<DeviceContext, T>& blas, const int n,
+                        const T* x, T* y) {
+  T scalar = x[0];
+  // max
+  for (int i = 1; i < n; ++i) {
+    scalar = scalar < x[i] ? x[i] : scalar;
+  }
+
+  // sub
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] - scalar;
+  }
+
+  // exp
+  blas.VEXP(n, y, y);
+
+  // sum
+  scalar = T(0);
+  for (int i = 0; i < n; ++i) {
+    scalar += y[i];
+  }
+
+  // scale
+  blas.SCAL(n, static_cast<T>(1) / scalar, y);
+}
+
+template <typename T>
+class AttentionLSTMKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using DeviceContext = paddle::platform::CPUDeviceContext;
+
+    auto* x = ctx.Input<LoDTensor>("X");
+    auto* h0 = ctx.Input<Tensor>("H0");
+    auto* c0 = ctx.Input<Tensor>("C0");
+    auto* atten_w = ctx.Input<Tensor>("AttentionWeight");
+    auto* atten_b = ctx.Input<Tensor>("AttentionBias");
+    auto* atten_scalar = ctx.Input<Tensor>("AttentionScalar");
+    auto* atten_scalar_bias = ctx.Input<Tensor>("AttentionScalarBias");
+    auto* lstm_w = ctx.Input<Tensor>("LSTMWeight");
+    auto* lstm_b = ctx.Input<Tensor>("LSTMBias");
+
+    auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
+    auto* cell_out = ctx.Output<LoDTensor>("Cell");
+    auto* atted_x = ctx.Output<Tensor>("AttentionedX");
+    auto* fc_out = ctx.Output<Tensor>("AttentionFCOut");
+    auto* lstm_x = ctx.Output<Tensor>("LSTMX");
+    auto* lstm_out = ctx.Output<Tensor>("LSTMOUT");
+
+    // some shape should be reshape here since infershape can not get lod info
+    auto x_lod = x->lod();
+    const int N = x_lod[0].size() - 1;  // batch size
+    auto x_dims = x->dims();            // T x M
+    auto w_dims = lstm_w->dims();       // (D+M) x 4D
+    const int total_T = x_dims[0];
+    const int M = x_dims[1];      // x frame size
+    const int D = w_dims[1] / 4;  // gate frame size
+    const int D2 = D * 2;
+    const int D3 = D * 3;
+    const int D4 = w_dims[1];
+    int max_seq_len = x_lod[0][1];
+    for (int i = 1; i < N; ++i) {
+      int len = x_lod[0][i + 1] - x_lod[0][i];
+      max_seq_len = max_seq_len < len ? len : max_seq_len;
+    }
+    PADDLE_ENFORCE_EQ(x_lod.size(), 1, "Input(X)'s lod size must be 1.");
+    PADDLE_ENFORCE_EQ(c0->dims()[0], N, "C0 dims should be %d x %d.", N, D);
+    fc_out->Resize({max_seq_len, 1});
+
+    math::VecActivations<T> act_functor;
+    std::function<void(const int, const T *, T *)> act_gate, act_cell, act_cand;
+    act_gate = act_functor(ctx.Attr<std::string>("gate_activation"));
+    act_cell = act_functor(ctx.Attr<std::string>("cell_activation"));
+    act_cand = act_functor(ctx.Attr<std::string>("candidate_activation"));
+
+    const T* x_data = x->data<T>();
+    const T* h0_data = h0 ? h0->data<T>() : NULL;
+    const T* c0_data = c0->data<T>();
+    const T* lstm_w_data = lstm_w->data<T>();
+    const T* lstm_b_data = lstm_b->data<T>();
+    const T* atten_w_data = atten_w->data<T>();
+    const T* atten_b_data = atten_b ? atten_b->data<T>() : NULL;
+    const T* atten_scalar_data = atten_scalar ? atten_scalar->data<T>() : NULL;
+    const T* atten_scalar_bias_data =
+        atten_scalar_bias ? atten_scalar_bias->data<T>() : NULL;
+
+    T* hidden_out_data = hidden_out->mutable_data<T>(ctx.GetPlace());
+    T* cell_out_data = cell_out->mutable_data<T>(ctx.GetPlace());
+    T* atted_x_data = atted_x->mutable_data<T>(ctx.GetPlace());
+    T* fc_out_data = fc_out->mutable_data<T>(ctx.GetPlace());
+    T* lstm_x_data = lstm_x->mutable_data<T>(ctx.GetPlace());
+    T* lstm_out_data = lstm_out->mutable_data<T>(ctx.GetPlace());
+
+    // x(TxM) * fc (Mx1) part of atten_wgt(M+D)x1
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    math::FCCompute<DeviceContext, T>(blas, total_T, 1, M, x_data, atten_w_data,
+                                      atted_x_data, atten_b_data);
+
+    const T* cur_atten_x_data = atted_x_data;
+    const T* cur_x_data = x_data;
+    const T* prev_cell_data = NULL;
+    const T* prev_hidden_data = NULL;
+    T* cur_cell_out_data = cell_out_data;
+    T* cur_hidden_out_data = hidden_out_data;
+    for (int i = 0; i < N; ++i) {
+      int seq_len = x_lod[0][i + 1] - x_lod[0][i];
+      prev_cell_data = c0_data + i * D;
+      prev_hidden_data = h0_data ? h0_data + i * D : NULL;
+      for (int step = 0; step < seq_len; ++step) {
+        /// 1. compute attention vector
+        // 1a. prev_cell(1xD) * fc(D) rest part of atten_wgt
+        T prev_cell_bias = blas.DOT(D, prev_cell_data, atten_w_data + M);
+        // 1b. add cell bias and relu
+        bias_relu<T>(seq_len, cur_atten_x_data, &prev_cell_bias, fc_out_data);
+        // 1c. fc scalar
+        if (atten_scalar_data) {
+          blas.SCAL(seq_len, *atten_scalar_data, fc_out_data);
+          bias_relu<T>(seq_len, fc_out_data, atten_scalar_bias_data,
+                       fc_out_data);
+        }
+        // 1d. softmax
+        vec_softmax<DeviceContext, T>(blas, seq_len, fc_out_data, fc_out_data);
+        // mul x(seq_len*M) and sum pool
+        math::FCCompute<DeviceContext, T>(blas, 1, M, seq_len, fc_out_data,
+                                          cur_x_data, lstm_x_data);
+
+        /// 2. compute LSTM step
+        // lstm weight : concat[forget , input , output , tilde]
+        // shape : (D + M) x (4 * D)
+        // fc inputX(1xM) * weightX(M*(4D))  => 1 x 4D
+        blas.MatMul(1, D4, M, lstm_x_data, lstm_w_data + D * D4, lstm_out_data);
+        if (prev_hidden_data) {
+          blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D4, D, static_cast<T>(1),
+                    prev_hidden_data, D, lstm_w_data, D4, static_cast<T>(1),
+                    lstm_out_data, D4);
+        }
+        // since input is 1xM, so can use add bias
+        blas.VADD(D4, lstm_b_data, lstm_out_data, lstm_out_data);
+
+        // gate act: sigmoid
+        act_gate(D3, lstm_out_data, lstm_out_data);
+        // candicate act: tanh
+        act_cand(D, lstm_out_data + D3, lstm_out_data + D3);
+
+        // a = forget * prev_cell
+        blas.VMUL(D, lstm_out_data, prev_cell_data, lstm_out_data);
+
+        // b = input * tilde
+        blas.VMUL(D, lstm_out_data + D, lstm_out_data + D3, lstm_out_data + D);
+
+        // cell_out = a + b
+        blas.VADD(D, lstm_out_data, lstm_out_data + D, cur_cell_out_data);
+
+        // state act tanh(cell_out) * output_gate
+        act_cell(D, cur_cell_out_data, lstm_out_data);
+        blas.VMUL(D, lstm_out_data, lstm_out_data + D2, cur_hidden_out_data);
+
+        prev_hidden_data = cur_hidden_out_data;
+        prev_cell_data = cur_cell_out_data;
+        cur_cell_out_data = cur_cell_out_data + D;
+        cur_hidden_out_data = cur_hidden_out_data + D;
+      }
+      cur_x_data = cur_x_data + seq_len * M;
+      cur_atten_x_data = cur_atten_x_data + seq_len;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(attention_lstm, ops::AttentionLSTMOp,
+                  ops::AttentionLSTMOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+
+REGISTER_OP_CPU_KERNEL(attention_lstm, ops::AttentionLSTMKernel<float>,
+                       ops::AttentionLSTMKernel<double>);

paddle/fluid/operators/attention_lstm_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+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/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
+
+class AttentionLSTMOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class AttentionLSTMOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override;
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/fusion_lstm_op.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-// #include <string>
 #include "paddle/fluid/framework/op_registry.h"
 
 namespace paddle {

paddle/fluid/operators/math/blas.h

@@ -90,6 +90,11 @@ class Blas {
   void GEMM(bool transA, bool transB, int M, int N, int K, T alpha, const T* A,
             int lda, const T* B, int ldb, T beta, T* C, int ldc) const;
 
+  template <typename T>
+  void GEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N, int K,
+            T alpha, const T* A, int lda, const T* B, int ldb, T beta, T* C,
+            int ldc) const;
+
 #ifdef PADDLE_WITH_MKLML
   template <typename T>
   T* GEMM_ALLOC(const CBLAS_IDENTIFIER id, const int M, const int N,
@@ -109,6 +114,10 @@ class Blas {
   void GEMM_FREE(T* data) const;
 #endif
 
+  template <typename T>
+  void MatMul(const int M, const int N, const int K, const T* A, const T* B,
+              T* C) const;
+
   template <typename T>
   void MatMul(const framework::Tensor& mat_a, bool trans_a,
               const framework::Tensor& mat_b, bool trans_b, T alpha,
@@ -140,10 +149,19 @@ class Blas {
   template <typename T>
   void VCOPY(int n, const T* x, T* y) const;
 
+  template <typename T>
+  void VEXP(int n, const T* x, T* y) const;
+
   template <typename T>
   void GEMV(bool trans_a, int M, int N, T alpha, const T* A, const T* B, T beta,
             T* C) const;
 
+  template <typename T>
+  T DOT(int n, const T* x, const T* y) const;
+
+  template <typename T>
+  void SCAL(int n, const T a, T* x) const;
+
   template <typename T>
   void BatchedGEMM(CBLAS_TRANSPOSE transA, CBLAS_TRANSPOSE transB, int M, int N,
                    int K, T alpha, const T* A, const T* B, T beta, T* C,
@@ -215,11 +233,26 @@ class BlasT : private Blas<DeviceContext> {
     Base()->template VCOPY<T>(args...);
   }
 
+  template <typename... ARGS>
+  void VEXP(ARGS... args) const {
+    Base()->template VEXP<T>(args...);
+  }
+
   template <typename... ARGS>
   void GEMV(ARGS... args) const {
     Base()->template GEMV<T>(args...);
   }
 
+  template <typename... ARGS>
+  T DOT(ARGS... args) const {
+    return Base()->template DOT<T>(args...);
+  }
+
+  template <typename... ARGS>
+  void SCAL(ARGS... args) const {
+    Base()->template SCAL<T>(args...);
+  }
+
   template <typename... ARGS>
   void BatchedGEMM(ARGS... args) const {
     Base()->template BatchedGEMM<T>(args...);

paddle/fluid/operators/math/blas_impl.h

@@ -73,6 +73,16 @@ struct CBlas<float> {
     platform::dynload::cblas_sgemv(args...);
   }
 
+  template <typename... ARGS>
+  static float DOT(ARGS... args) {
+    return platform::dynload::cblas_sdot(args...);
+  }
+
+  template <typename... ARGS>
+  static void SCAL(ARGS... args) {
+    platform::dynload::cblas_sscal(args...);
+  }
+
   template <typename... ARGS>
   static void GEMM_BATCH(ARGS... args) {
     platform::dynload::cblas_sgemm_batch(args...);
@@ -87,6 +97,11 @@ struct CBlas<float> {
   static void VMUL(ARGS... args) {
     platform::dynload::vsMul(args...);
   }
+
+  template <typename... ARGS>
+  static void VEXP(ARGS... args) {
+    platform::dynload::vsExp(args...);
+  }
 };
 
 template <>
@@ -138,6 +153,16 @@ struct CBlas<double> {
     platform::dynload::cblas_dgemv(args...);
   }
 
+  template <typename... ARGS>
+  static double DOT(ARGS... args) {
+    return platform::dynload::cblas_ddot(args...);
+  }
+
+  template <typename... ARGS>
+  static void SCAL(ARGS... args) {
+    platform::dynload::cblas_dscal(args...);
+  }
+
   template <typename... ARGS>
   static void GEMM_BATCH(ARGS... args) {
     platform::dynload::cblas_dgemm_batch(args...);
@@ -152,6 +177,11 @@ struct CBlas<double> {
   static void VMUL(ARGS... args) {
     platform::dynload::vdMul(args...);
   }
+
+  template <typename... ARGS>
+  static void VEXP(ARGS... args) {
+    platform::dynload::vdExp(args...);
+  }
 };
 
 #else
@@ -210,6 +240,9 @@ struct CBlas<platform::float16> {
     PADDLE_THROW("float16 SMM_GEMM not supported on CPU");
   }
   static void VMUL(...) { PADDLE_THROW("float16 VMUL not supported on CPU"); }
+  static void VEXP(...) { PADDLE_THROW("float16 VEXP not supported on CPU"); }
+  static void DOT(...) { PADDLE_THROW("float16 DOT not supported on CPU"); };
+  static void SCAL(...) { PADDLE_THROW("float16 SCAL not supported on CPU"); };
 #ifdef PADDLE_WITH_MKLML
   static void GEMM_BATCH(...) {
     PADDLE_THROW("float16 GEMM_BATCH not supported on CPU");
@@ -217,64 +250,6 @@ struct CBlas<platform::float16> {
 #endif
 };
 
-template <typename T>
-inline bool UseXSMM(const int &m, const int &n, const int &k, bool transa,
-                    bool transb, const T &alpha, const T &beta) {
-#ifdef PADDLE_WITH_LIBXSMM
-  // Refer to https://github.com/hfp/libxsmm/blob/master/README.md
-  // But the threshold is custom
-  constexpr int LIBXSMM_THRESHOLD = 20 * 20 * 20;
-  if (m * n * k > LIBXSMM_THRESHOLD || transa || transb ||
-      std::abs<T>(alpha - static_cast<T>(1) >
-                  std::numeric_limits<T>::epsilon()) ||
-      std::abs<T>(beta) > std::numeric_limits<T>::epsilon()) {
-    return false;
-  } else {
-    return true;
-  }
-#endif
-  return false;
-}
-
-template <>
-inline bool UseXSMM<platform::float16>(const int &m, const int &n, const int &k,
-                                       bool transa, bool transb,
-                                       const platform::float16 &alpha,
-                                       const platform::float16 &beta) {
-  return false;
-}
-
-template <typename T>
-inline void GEMM_WARP(CBLAS_ORDER order, CBLAS_TRANSPOSE transA,
-                      CBLAS_TRANSPOSE transB, int M, int N, int K, T alpha,
-                      const T *A, int lda, const T *B, int ldb, T beta, T *C,
-                      int ldc) {
-#ifdef PADDLE_WITH_LIBXSMM
-  if (UseXSMM<T>(M, N, K, transA != CblasNoTrans, transB != CblasNoTrans, alpha,
-                 beta)) {
-    // Note: SMM use ColMajor
-    const char transa = 'N';
-    const char transb = 'N';
-    CBlas<T>::SMM_GEMM(&transa, &transb, &N, &M, &K, &alpha, B, &ldb, A, &lda,
-                       &beta, C, &ldc);
-    return;
-  }
-#endif
-
-#ifdef PADDLE_MKL_SPLIT_GEMM
-  constexpr int bs = 2;
-  if (M % bs == 0 && transA == CblasNoTrans && transB == CblasNoTrans) {
-    for (int off = 0; off < M; off += bs) {
-      CBlas<T>::GEMM(CblasRowMajor, CblasNoTrans, CblasNoTrans, bs, N, K, alpha,
-                     A + off * lda, lda, B, ldb, beta, C + off * ldb, ldc);
-    }
-    return;
-  }
-#endif
-  CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
-                 beta, C, ldc);
-}
-
 #ifdef PADDLE_WITH_MKLML
 template <>
 template <typename T>
@@ -319,7 +294,7 @@ void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
-  GEMM_WARP<T>(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
+  CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
                  beta, C, ldc);
 }
 
@@ -329,11 +304,22 @@ void Blas<platform::CPUDeviceContext>::GEMM(bool transA, bool transB, int M,
                                             int N, int K, T alpha, const T *A,
                                             int lda, const T *B, int ldb,
                                             T beta, T *C, int ldc) const {
-  GEMM_WARP<T>(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
+  CBlas<T>::GEMM(CblasRowMajor, transA == false ? CblasNoTrans : CblasTrans,
                  transB == false ? CblasNoTrans : CblasTrans, M, N, K, alpha, A,
                  lda, B, ldb, beta, C, ldc);
 }
 
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
+                                            CBLAS_TRANSPOSE transB, int M,
+                                            int N, int K, T alpha, const T *A,
+                                            int lda, const T *B, int ldb,
+                                            T beta, T *C, int ldc) const {
+  CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
+                 beta, C, ldc);
+}
+
 template <typename DeviceContext>
 template <typename T>
 void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a, bool trans_a,
@@ -399,6 +385,47 @@ void Blas<platform::CPUDeviceContext>::VMUL(int n, const T *x, const T *y,
 #endif
 }
 
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::VEXP(int n, const T *x, T *y) const {
+#ifdef PADDLE_WITH_MKLML
+  CBlas<T>::VEXP(n, x, y);
+#else
+  // try to find if openblas support vexp
+  for (int i = 0; i < n; ++i) {
+    y[i] = std::exp(x[i]);
+  }
+#endif
+}
+
+template <>
+template <typename T>
+T Blas<platform::CPUDeviceContext>::DOT(int n, const T *x, const T *y) const {
+#ifdef PADDLE_WITH_MKLML
+  return CBlas<T>::DOT(n, x, 1, y, 1);
+#else
+  // try to find if openblas support cblas_dot
+  T sum = 0;
+  for (int i = 0; i < n; ++i) {
+    sum += x[i] * y[i];
+  }
+  return sum;
+#endif
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::SCAL(int n, const T a, T *x) const {
+#ifdef PADDLE_WITH_MKLML
+  CBlas<T>::SCAL(n, a, x, 1);
+#else
+  // try to find if openblas support cblas_scal
+  for (int i = 0; i < n; ++i) {
+    x[i] = a * x[i];
+  }
+#endif
+}
+
 template <>
 template <typename T>
 void Blas<platform::CPUDeviceContext>::GEMV(bool trans_a, int M, int N, T alpha,
@@ -440,6 +467,42 @@ void Blas<platform::CPUDeviceContext>::BatchedGEMM(
 #endif
 }
 
+template <typename DeviceContext>
+template <typename T>
+void Blas<DeviceContext>::MatMul(const int M, const int N, const int K,
+                                 const T *A, const T *B, T *C) const {
+  this->template GEMM<T>(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K,
+                         static_cast<T>(1), A, K, B, N, static_cast<T>(0), C,
+                         N);
+}
+
+template <>
+template <typename T>
+void Blas<platform::CPUDeviceContext>::MatMul(const int M, const int N,
+                                              const int K, const T *A,
+                                              const T *B, T *C) const {
+#ifdef PADDLE_WITH_LIBXSMM
+  // Refer to https://github.com/hfp/libxsmm/blob/master/README.md
+  // But the threshold is custom constexpr int LIBXSMM_THRESHOLD = 20 * 20 * 20;
+
+  // Since the matrix is very small,
+  // so the unit of calculation is already very fast,
+  // and the if( M*N*K < LIBXSMM_THRESHOLD) would be overhead,
+  // use xsmm directly.
+  // Note: SMM use ColMajor
+  const char transa = 'N';
+  const char transb = 'N';
+  const T alpha = static_cast<T>(1);
+  const T beta = static_cast<T>(0);
+  CBlas<T>::SMM_GEMM(&transa, &transb, &N, &M, &K, &alpha, B, &N, A, &K, &beta,
+                     C, &N);
+  return;
+#endif
+
+  CBlas<T>::GEMM(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K,
+                 static_cast<T>(1), A, K, B, N, static_cast<T>(0), C, N);
+}
+
 template <typename DeviceContext>
 template <typename T>
 void Blas<DeviceContext>::MatMul(const framework::Tensor &mat_a,

paddle/fluid/operators/math/cpu_vec.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+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 <string>
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+#define SIGMOID_THRESHOLD_MIN -40.0
+#define SIGMOID_THRESHOLD_MAX 13.0
+#define EXP_MAX_INPUT 40.0
+
+template <typename T>
+inline T sigmoid(T x) {
+  return 1. / (1. + exp(-x));
+}
+
+template <typename T>
+inline T tanh(T x) {
+  return 2. * sigmoid(2. * x) - 1.;
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_identity(const int n, const T* x, T* y) {
+  // do nothing
+  return;
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_sigmoid(const int n, const T* x, T* y) {
+  const T min = SIGMOID_THRESHOLD_MIN;
+  const T max = SIGMOID_THRESHOLD_MAX;
+  for (int i = 0; i < n; ++i) {
+    T tmp = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
+    y[i] = 1.0 / (1.0 + std::exp(-tmp));
+  }
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_tanh(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = tanh<T>(x[i]);
+  }
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+inline void vec_relu(const int n, const T* x, T* y) {
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0 ? x[i] : 0;
+  }
+}
+
+template <>
+inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
+                                                 float* y) {
+  // TODO(TJ): complete me
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0 ? x[i] : 0;
+  }
+}
+
+template <>
+inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
+                                                float* y) {
+  // TODO(TJ): complete me
+  for (int i = 0; i < n; ++i) {
+    y[i] = x[i] > 0 ? x[i] : 0;
+  }
+}
+
+template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
+class VecActivations {
+ public:
+  std::function<void(const int, const T*, T*)> operator()(
+      const std::string& type) {
+    if (type == "sigmoid") {
+      return vec_sigmoid<T, isa>;
+    } else if (type == "relu") {
+      return vec_relu<T, isa>;
+    } else if (type == "tanh") {
+      return vec_tanh<T, isa>;
+    } else if (type == "identity" || type == "") {
+      return vec_identity<T, isa>;
+    }
+    PADDLE_THROW("Not support type %s.", type);
+  }
+};
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/math/fc_compute.h

@@ -25,17 +25,25 @@ namespace math {
 template <typename DeviceContext, typename T>
 inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
                       const int N, const int K, const T* X, const T* W, T* Y,
-                      const T* B = NULL) {
-  blas.GEMM(CblasNoTrans, CblasNoTrans, M, N, K, static_cast<T>(1), X, W,
-            static_cast<T>(0), Y);
-  if (B) {
+                      const T* B = NULL, bool relu = false) {
+  blas.MatMul(M, N, K, X, W, Y);
+  if (B == NULL) {
+    return;
+  }
+
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for if (FLAGS_paddle_num_threads > 1)
 #endif
   for (int i = 0; i < M; i++) {
     blas.AXPY(N, static_cast<T>(1), B, Y + i * N);
   }
+
+  if (!relu) {
+    return;
   }
+
+  // TODO(TJ): fuse relu
+  LOG(FATAL) << "Not implemented!";
 }
 
 }  // namespace math

paddle/fluid/platform/cpu_info.cc

@@ -103,15 +103,16 @@ size_t CUDAPinnedMaxChunkSize() {
   return CUDAPinnedMaxAllocSize() / 256;
 }
 
-#ifdef PADDLE_WITH_XBYAK
 namespace jit {
-
+#ifdef PADDLE_WITH_XBYAK
 static Xbyak::util::Cpu cpu;
 bool MayIUse(const cpu_isa_t cpu_isa) {
   using namespace Xbyak::util;  // NOLINT
   switch (cpu_isa) {
     case sse42:
       return cpu.has(Cpu::tSSE42);
+    case avx:
+      return cpu.has(Cpu::tAVX);
     case avx2:
       return cpu.has(Cpu::tAVX2);
     case avx512_common:
@@ -134,8 +135,16 @@ bool MayIUse(const cpu_isa_t cpu_isa) {
   }
   return false;
 }
+#else
+bool MayIUse(const cpu_isa_t cpu_isa) {
+  if (cpu_isa == isa_any) {
+    return true;
+  } else {
+    return false;
+  }
+}
+#endif
 
 }  // namespace jit
-#endif
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/cpu_info.h

@@ -37,12 +37,11 @@ size_t CUDAPinnedMinChunkSize();
 //! Get the maximum chunk size for buddy allocator.
 size_t CUDAPinnedMaxChunkSize();
 
-#ifdef PADDLE_WITH_XBYAK
 namespace jit {
-
 typedef enum {
   isa_any,
   sse42,
+  avx,
   avx2,
   avx512_common,
   avx512_core,
@@ -55,7 +54,6 @@ typedef enum {
 inline bool MayIUse(const cpu_isa_t cpu_isa);
 
 }  // namespace jit
-#endif
 
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/dynload/mklml.h

@@ -66,10 +66,16 @@ extern void* mklml_dso_handle;
   __macro(cblas_dgemm_free);        \
   __macro(cblas_sgemm_batch);       \
   __macro(cblas_dgemm_batch);       \
+  __macro(cblas_sdot);              \
+  __macro(cblas_ddot);              \
+  __macro(cblas_sscal);             \
+  __macro(cblas_dscal);             \
   __macro(vsAdd);                   \
   __macro(vdAdd);                   \
   __macro(vsMul);                   \
   __macro(vdMul);                   \
+  __macro(vsExp);                   \
+  __macro(vdExp);                   \
   __macro(MKL_Set_Num_Threads)
 
 MKLML_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_MKLML_WRAP);

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

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+from op_test import OpTest
+from test_fusion_lstm_op import fc, ACTIVATION
+from test_softmax_op import stable_softmax
+
+
+def attention_lstm(
+        x,  # T x M
+        lod,  # 1 x N
+        h0,  # N x D
+        c0,  # N x D
+        fcws,  # (M+D) x 1, 1x1
+        fcbs,  # 1 x 1, 1x1
+        w,  # (M+D) x 4D
+        b,  # 1 x 4D
+        act_gate,
+        act_cell,
+        act_cand):
+
+    T = sum(lod[0])
+    N = len(lod[0])
+    M = x.shape[1]
+    D = b.shape[1] / 4
+    assert T == x.shape[0]
+    assert len(fcws) == len(fcbs)
+    hidden = []
+    cell = []
+
+    start_offset = 0
+    for bid in range(N):
+        seq_len = lod[0][bid]
+        xi = np.copy(x[start_offset:start_offset + seq_len, :]).reshape(seq_len,
+                                                                        M)
+        prev_cell = np.copy(c0[bid]).reshape([1, D])
+        prev_hidden = np.copy(h0[bid]).reshape([1, D])
+        for step in range(seq_len):
+            expanded_cell = np.repeat(prev_cell, seq_len, axis=0)
+            tmp = np.concatenate((xi, expanded_cell), axis=1)
+            assert tmp.shape[0] == seq_len
+            assert tmp.shape[1] == M + D
+            for fcid in range(len(fcbs)):
+                tmp = fc(tmp, fcws[fcid], fcbs[fcid])
+                tmp = ACTIVATION['relu'](tmp)
+            tmp = np.reshape(tmp, (1, seq_len))
+            tmp = stable_softmax(tmp).reshape(seq_len, 1)
+            lstmx = xi * tmp  # seq * M
+            lstmx = np.sum(lstmx.reshape(seq_len, M), axis=0).reshape([1, M])
+            lstmin = np.concatenate((prev_hidden, lstmx), axis=1)
+            lstmout = fc(lstmin, w, b).reshape([1, 4 * D])
+
+            g_f, g_i, g_o, cand = np.split(lstmout, 4, axis=1)
+            g_f = act_gate(g_f).reshape([1, D])
+            g_i = act_gate(g_i).reshape([1, D])
+            g_o = act_gate(g_o).reshape([1, D])
+            cand = act_cand(cand).reshape([1, D])
+
+            cell_t = (prev_cell * g_f) + (g_i * cand)
+            hidden_t = g_o * act_cell(cell_t)
+
+            hidden.append(hidden_t.flatten())
+            cell.append(cell_t.flatten())
+
+            prev_cell = cell_t.reshape([1, D])
+            prev_hidden = hidden_t.reshape([1, D])
+
+        start_offset += seq_len
+
+    hidden = np.array(hidden).astype('float32').reshape([T, D])
+    cell = np.array(cell).astype('float32').reshape([T, D])
+    return hidden, cell
+
+
+class TestAttentionLSTMOp(OpTest):
+    def set_conf(self):
+        pass
+
+    def setUp(self):
+        self.op_type = 'attention_lstm'
+        self.lod = [[3]]
+        self.M = 30
+        self.D = 15
+        self.has_initial_hidden = True
+        self.act_gate = 'sigmoid'
+        self.act_cell = 'tanh'
+        self.act_cand = 'tanh'
+        self.set_conf()
+
+        T = sum(self.lod[0])
+        bs = len(self.lod[0])
+
+        x = np.random.normal(size=(T, self.M)).astype('float32')
+        c0 = np.random.normal(size=(bs, self.D)).astype('float32')
+        if self.has_initial_hidden:
+            h0 = np.random.normal(size=(bs, self.D)).astype('float32')
+        else:
+            h0 = np.zeros((bs, self.D)).astype('float32')
+
+        fcw1 = np.random.normal(size=(self.M + self.D, 1)).astype('float32')
+        fcb1 = np.random.normal(size=(1, 1)).astype('float32')
+        fcw2 = np.random.normal(size=(1, 1)).astype('float32')
+        fcb2 = np.random.normal(size=(1, 1)).astype('float32')
+
+        # lstm weight and bias
+        w = np.random.normal(size=(self.M + self.D,
+                                   self.D * 4)).astype('float32')
+        b = np.random.normal(size=(1, self.D * 4)).astype('float32')
+
+        h, c = attention_lstm(x, self.lod, h0, c0, [fcw1, fcw2], [fcb1, fcb2],
+                              w, b, ACTIVATION[self.act_gate],
+                              ACTIVATION[self.act_cell],
+                              ACTIVATION[self.act_cand])
+
+        self.inputs = {
+            'X': (x, self.lod),
+            'C0': c0,
+            'AttentionWeight': fcw1,
+            'AttentionBias': fcb1,
+            'AttentionScalar': fcw2,
+            'AttentionScalarBias': fcb2,
+            'LSTMWeight': w,
+            'LSTMBias': b
+        }
+
+        if self.has_initial_hidden:
+            self.inputs['H0'] = h0
+
+        self.outputs = {
+            'Hidden': (h, self.lod),
+            'Cell': (c, self.lod),
+        }
+        self.attrs = {
+            'gate_activation': self.act_gate,
+            'cell_activation': self.act_cell,
+            'candidate_activation': self.act_cand
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAttentionOpNonInit(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.has_initial_hidden = False
+
+
+class TestAttentionOpAct(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.M = 3
+        self.D = 2
+        self.act_gate = 'relu'
+        self.act_cell = 'tanh'
+        self.act_cand = 'sigmoid'
+
+
+class TestAttentionOpMD1(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.M = 36
+        self.D = 8
+
+
+class TestAttentionOpMD2(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.M = 8
+        self.D = 8
+
+
+class TestAttentionOpMD3(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.M = 15
+        self.D = 30
+
+
+class TestAttentionOpBS1(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.lod = [[5]]
+        self.M = 16
+        self.D = 32
+
+
+class TestAttentionOpBS2(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.lod = [[3, 6]]
+
+
+class TestAttentionOpBS5(TestAttentionLSTMOp):
+    def set_conf(self):
+        self.lod = [[3, 2, 4, 7, 5]]
+
+
+if __name__ == '__main__':
+    unittest.main()