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

paddle/fluid/API.spec

@@ -187,6 +187,7 @@ paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=Non
 paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
 paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
+paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'dropout_prob', 'input_size', 'hidden_size', 'num_layers', 'is_bidirec', 'dtype', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(False, 'float32', False, None, None, False, 0))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)

paddle/fluid/operators/lstm_cudnn_op.cc

+/* 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. */
+
+#include "paddle/fluid/operators/lstm_cudnn_op.h"
+#include <string>
+
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/cudnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+class CudnnLSTMOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("W"),
+                   "Input(Weight) of LSTM should not be null.");
+
+    PADDLE_ENFORCE(ctx->HasInput("InitH"),
+                   "Input(init_h) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("InitC"),
+                   "Input(init_c) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Cache"),
+                   "Input(Cache) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("last_h"),
+                   "Output(last_h) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("last_c"),
+                   "Output(last_c) of LSTM should not be null.");
+
+    auto in_dims = ctx->GetInputDim("Input");
+    PADDLE_ENFORCE_EQ(in_dims.size(), 3, "Input(X)'s rank must be 3.");
+
+    ctx->SetOutputDim("Out", ctx->GetInputDim("Input"));
+    ctx->SetOutputDim("last_h", ctx->GetInputDim("InitH"));
+    ctx->SetOutputDim("last_c", ctx->GetInputDim("InitC"));
+  }
+};
+
+class CudnnLSTMOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput(
+        "Input",
+        "(Tensor) RNN input tensor, which support variable-time length input "
+        "sequence."
+        "The shape of the Tensor MUST be ( seq_len * batch_size * input_size)"
+        "seq_len is the total time step in this mini-batch (CAN be change in "
+        "different batch)"
+        "batch_size is the instance number of this batch"
+        "input_size is the hidden size of the input."
+        "input_hidden_size and the hidden_size in the next may not be same");
+    AddInput("InitH",
+             "(Tensor) the initial hidden state of the LSTM"
+             "input. This is a tensor with shape (num_layers x batch_size x "
+             "hidden_size)"
+             "and When is_bidirec is True, the shape will be (num_layers*2 x "
+             "batch_size x hidden_size)");
+    AddInput("InitC",
+             "(Tensor) the initial cell state of the LSTm "
+             "input. This is a tensor with shape (num_layers x batch_size x "
+             "hidden_size)"
+             "and When is_bidirec is True, the shape will be (num_layers*2 x "
+             "batch_size x hidden_size)");
+    AddInput("W",
+             "(Tensor) the learnable hidden-hidden weights."
+             " The shape is (N), where N is total weight size of the LSTM. "
+             " cudnn concatenate all the weight to one Tensor");
+    AddInput("Cache",
+             "The cache of dropout op, a RAW type variable including random "
+             "number generator states and some descriptors, which is used in "
+             "cudnn kernel.")
+        .AsDispensable();
+    AddOutput("Out",
+              "(Tensor) the hidden state of LSTM operator. "
+              "The shape is ( seq_len x batch_size x hidden_size) if "
+              "is_bidirec is False"
+              "and When is_bidirec is True, the shape will be ( seq_len x "
+              "batch_size x hidden_size * 2) ");
+    AddOutput("last_h",
+              "(Tensor) the hidden state of the last step. "
+              "The shape is ( num_layers x batch_size x hidden_size) if "
+              "is_bidirec is False"
+              "and When is_bidirec is True, the shape will be (num_layers*2 x "
+              "batch_size x hidden_size)");
+    AddOutput("last_c",
+              "(Tensor) the cell state of the last step"
+              "The shape is ( num_layers x batch_size x hidden_size) if "
+              "is_bidirec is False"
+              "and When is_bidirect is True, the shape will be (num_layers*2 x "
+              "batch_size x hidden_size*2)");
+    AddAttr<int>("max_len",
+                 "max length of the LSTM op"
+                 "the first dim of the Input can NOT be greater than max_len")
+        .SetDefault(20);
+    AddAttr<float>(
+        "dropout_prob",
+        "dropout prob of the dropout op"
+        "the dropout ONLY work between lstm layers, not between time steps"
+        "There is no dropout work on the Out tensor")
+        .SetDefault(0.0);
+    AddAttr<bool>("is_bidirec",
+                  "is_bidirec"
+                  "if it is bidirection rnn"
+                  "The will affect the shape of the Out, last_h, and last_c")
+        .SetDefault(false);
+    AddAttr<int>("input_size", "input size ot the Input Tensor").SetDefault(10);
+    AddAttr<int>("hidden_size", "hidden size of the LSTM").SetDefault(100);
+    AddAttr<int>("num_layers", "the total layer number of the LSTM")
+        .SetDefault(1);
+    AddAttr<bool>("is_test", "True if in test phase.").SetDefault(false);
+    AddAttr<int>("seed", "seed to used if fix_seed is True").SetDefault(-1);
+    AddComment(R"DOC(
+CUDNN LSTM implementation
+
+A four-gate Long Short-Term Memory network with no peephole connections.
+In the forward pass the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1, 
+the cell input ct-1 and the previous layer input xt given matrices W, R and biases bW, bR from the following equations:
+
+$$ i_t = sigmoid(W_{ix}x_{t} + W_{ih}h_{t-1} + bx_i + bh_i) $$
+
+$$ f_t = sigmoid(W_{fx}x_{t} + W_{fh}h_{t-1} + bx_f + bh_f) $$
+
+$$ o_t = sigmoid(W_{ox}x_{t} + W_{oh}h_{t-1} + bx_o + bh_o) $$
+
+$$ \\tilde{c_t} = tanh(W_{cx}x_t + W_{ch}h_{t-1} + bx_c + bh_c) $$
+
+$$ c_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} $$
+
+$$ h_t = o_t \\odot tanh(c_t) $$
+
+- W terms denote weight matrices (e.g. $W_{ix}$ is the matrix
+  of weights from the input gate to the input)
+- The b terms denote bias vectors ($bx_i$ and $bh_i$ are the input gate bias vector).
+- sigmoid is the logistic sigmoid function.
+- $i, f, o$ and $c$ are the input gate, forget gate, output gate,
+  and cell activation vectors, respectively, all of which have the same size as
+  the cell output activation vector $h$.
+- The $\odot$ is the element-wise product of the vectors.
+- `tanh` is the activation functions.
+- $\tilde{c_t}$ is also called candidate hidden state,
+  which is computed based on the current input and the previous hidden state.
+
+Where sigmoid is the sigmoid operator: sigmoid(x) = 1 / (1 + e^-x), * represents a point-wise multiplication, 
+X represensts a matrix multiplication
+
+
+)DOC");
+  }
+};
+
+class CudnnLSTMGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("last_h"),
+                   "Input(last_h) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("last_c"),
+                   "Input(last_c) of LSTM should not be null.");
+
+    PADDLE_ENFORCE(ctx->HasInput("Cache"),
+                   "Input(last_c) of LSTM should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("InitH"),
+                   "Input(init_h) of LSTM should not be null.");
+
+    PADDLE_ENFORCE(ctx->HasInput("InitC"),
+                   "Input(init_c) of LSTM should not be null.");
+
+    auto SetOutGradDim = [&ctx](const std::string& name) {
+      auto g_name = framework::GradVarName(name);
+      if (ctx->HasOutput(g_name)) {
+        ctx->SetOutputDim(g_name, ctx->GetInputDim(name));
+      }
+    };
+
+    SetOutGradDim("Input");
+    SetOutGradDim("W");
+    SetOutGradDim("InitH");
+    SetOutGradDim("InitC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(cudnn_lstm, ops::CudnnLSTMOp, ops::CudnnLSTMOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(cudnn_lstm_grad, ops::CudnnLSTMGradOp);
+
+REGISTER_OP_CPU_KERNEL(
+    cudnn_lstm,
+    ops::CudnnLSTMKernel<paddle::platform::CPUDeviceContext, float>);
+
+REGISTER_OP_CPU_KERNEL(
+    cudnn_lstm_grad,
+    ops::CudnnLSTMGradKernel<paddle::platform::CPUDeviceContext, float>);

paddle/fluid/operators/lstm_cudnn_op.cu.cc

+/* 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. */
+
+#include "paddle/fluid/operators/lstm_cudnn_op.h"
+#include "paddle/fluid/platform/cudnn_helper.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
+
+struct CudnnRNNCache {
+  CudnnRNNCache() {
+    x_desc_ = NULL;
+    y_desc_ = NULL;
+    dx_desc_ = NULL;
+    dy_desc_ = NULL;
+  }
+  ~CudnnRNNCache() { release(); }
+
+  cudnnRNNDescriptor_t rnn_desc_;
+  cudnnTensorDescriptor_t *x_desc_;
+  cudnnTensorDescriptor_t *y_desc_;
+  cudnnTensorDescriptor_t *dx_desc_;
+  cudnnTensorDescriptor_t *dy_desc_;
+
+  cudnnTensorDescriptor_t hx_desc_;
+  cudnnTensorDescriptor_t cx_desc_;
+  cudnnTensorDescriptor_t hy_desc_;
+  cudnnTensorDescriptor_t cy_desc_;
+
+  cudnnTensorDescriptor_t dhx_desc_;
+  cudnnTensorDescriptor_t dcx_desc_;
+  cudnnTensorDescriptor_t dhy_desc_;
+  cudnnTensorDescriptor_t dcy_desc_;
+
+  cudnnTensorDescriptor_t output_x_desc_;
+  cudnnTensorDescriptor_t output_y_desc_;
+
+  cudnnDropoutDescriptor_t dropout_desc_;
+
+  size_t weights_size_;
+  cudnnFilterDescriptor_t w_desc_;
+  cudnnFilterDescriptor_t dw_desc_;
+
+  size_t workspace_size_;
+  size_t reserve_size_;
+  Tensor reserve_data_;
+  Tensor workspace_data_;
+
+  Tensor dropout_state_;
+
+  size_t max_length_;
+
+  float dropout_prob_;
+  bool is_bidirec_;
+
+  int batch_size_;
+  int input_size_;
+  int hidden_size_;
+  int num_layers_;
+  int seed_;
+
+  void init(cudnnHandle_t handle, const framework::ExecutionContext &ctx,
+            size_t max_len, int batch_size, int input_size, int hidden_size,
+            int num_layers, float dropout_prob, bool is_bidirec, int seed,
+            int weight_numel) {
+    max_length_ = max_len;
+    batch_size_ = batch_size;
+    input_size_ = input_size;
+    hidden_size_ = hidden_size;
+    num_layers_ = num_layers;
+    dropout_prob_ = dropout_prob;
+    is_bidirec_ = is_bidirec;
+    seed_ = seed;
+
+    x_desc_ = new cudnnTensorDescriptor_t[max_length_];
+    y_desc_ = new cudnnTensorDescriptor_t[max_length_];
+    dx_desc_ = new cudnnTensorDescriptor_t[max_length_];
+    dy_desc_ = new cudnnTensorDescriptor_t[max_length_];
+    int dim_a[3];
+    int stride_a[3];
+
+    for (size_t i = 0; i < max_length_; ++i) {
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnCreateTensorDescriptor(&x_desc_[i]));
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnCreateTensorDescriptor(&y_desc_[i]));
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnCreateTensorDescriptor(&dx_desc_[i]));
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnCreateTensorDescriptor(&dy_desc_[i]));
+      dim_a[0] = batch_size_;
+      dim_a[1] = input_size_;
+      dim_a[2] = 1;
+
+      stride_a[0] = dim_a[2] * dim_a[1];
+      stride_a[1] = dim_a[2];
+      stride_a[2] = 1;
+      CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+          x_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+      CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+          dx_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+
+      dim_a[0] = batch_size_;
+      dim_a[1] = is_bidirec_ ? hidden_size_ * 2 : hidden_size_;
+      dim_a[2] = 1;
+
+      stride_a[0] = dim_a[2] * dim_a[1];
+      stride_a[1] = dim_a[2];
+      stride_a[2] = 1;
+
+      CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+          y_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+      CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+          dy_desc_[i], CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    }
+
+    dim_a[0] = num_layers_ * (is_bidirec_ ? 2 : 1);
+    dim_a[1] = batch_size_;
+    dim_a[2] = hidden_size_;
+
+    stride_a[0] = dim_a[2] * dim_a[1];
+    stride_a[1] = dim_a[2];
+    stride_a[2] = 1;
+
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&hx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&cx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&hy_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&cy_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dhx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dcx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dhy_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateTensorDescriptor(&dcy_desc_));
+
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        hx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        cx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        hy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        cy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        dhx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        dcx_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        dhy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
+        dcy_desc_, CUDNN_DATA_FLOAT, 3, dim_a, stride_a));
+
+    CUDNN_ENFORCE(
+        platform::dynload::cudnnCreateDropoutDescriptor(&dropout_desc_));
+
+    size_t state_size;
+    CUDNN_ENFORCE(
+        platform::dynload::cudnnDropoutGetStatesSize(handle, &state_size);
+        dropout_state_.Resize({static_cast<int64_t>(state_size)}));
+    auto *dropout_state_data =
+        dropout_state_.mutable_data<uint8_t>(ctx.GetPlace());
+    CUDNN_ENFORCE(platform::dynload::cudnnSetDropoutDescriptor(
+        dropout_desc_, handle, dropout_prob_, dropout_state_data, state_size,
+        seed_));
+
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateRNNDescriptor(&rnn_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetRNNDescriptor_v6(
+        handle, rnn_desc_, hidden_size_, num_layers_, dropout_desc_,
+        CUDNN_LINEAR_INPUT,
+        is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, CUDNN_LSTM,
+        CUDNN_RNN_ALGO_STANDARD, CUDNN_DATA_FLOAT));
+
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateFilterDescriptor(&w_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnCreateFilterDescriptor(&dw_desc_));
+
+    CUDNN_ENFORCE(platform::dynload::cudnnGetRNNParamsSize(
+        handle, rnn_desc_, x_desc_[0], &weights_size_, CUDNN_DATA_FLOAT));
+
+    PADDLE_ENFORCE_EQ(weights_size_, sizeof(float) * weight_numel,
+                      "cudnn lstm weight size should be SAME");
+    int dim_w[3];
+    dim_w[0] = weights_size_ / sizeof(float);
+    dim_w[1] = 1;
+    dim_w[2] = 1;
+    CUDNN_ENFORCE(platform::dynload::cudnnSetFilterNdDescriptor(
+        w_desc_, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 3, dim_w));
+    CUDNN_ENFORCE(platform::dynload::cudnnSetFilterNdDescriptor(
+        dw_desc_, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 3, dim_w));
+
+    CUDNN_ENFORCE(platform::dynload::cudnnGetRNNWorkspaceSize(
+        handle, rnn_desc_, max_length_, x_desc_, &workspace_size_));
+    CUDNN_ENFORCE(platform::dynload::cudnnGetRNNTrainingReserveSize(
+        handle, rnn_desc_, max_length_, x_desc_, &reserve_size_));
+
+    reserve_data_.Resize({static_cast<int64_t>(reserve_size_)});
+    reserve_data_.mutable_data<uint8_t>(ctx.GetPlace());
+
+    workspace_data_.Resize({static_cast<int64_t>(workspace_size_)});
+    workspace_data_.mutable_data<uint8_t>(ctx.GetPlace());
+  }
+
+  void release() {
+    for (size_t i = 0; i < max_length_; ++i) {
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnDestroyTensorDescriptor(x_desc_[i]));
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnDestroyTensorDescriptor(y_desc_[i]));
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnDestroyTensorDescriptor(dx_desc_[i]));
+      CUDNN_ENFORCE(
+          platform::dynload::cudnnDestroyTensorDescriptor(dy_desc_[i]));
+    }
+
+    delete[] x_desc_;
+    delete[] y_desc_;
+    delete[] dx_desc_;
+    delete[] dy_desc_;
+
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(hx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(cx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(hy_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(cy_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dhx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dcx_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dhy_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(dcy_desc_));
+
+    CUDNN_ENFORCE(
+        platform::dynload::cudnnDestroyDropoutDescriptor(dropout_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyRNNDescriptor(rnn_desc_));
+
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyFilterDescriptor(w_desc_));
+    CUDNN_ENFORCE(platform::dynload::cudnnDestroyFilterDescriptor(dw_desc_));
+  }
+};
+
+template <typename DeviceContext, typename T>
+class CudnnLSTMGPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    const Tensor *x = ctx.Input<Tensor>("Input");
+    const Tensor *init_h = ctx.Input<Tensor>("InitH");
+    const Tensor *init_c = ctx.Input<Tensor>("InitC");
+
+    auto w = ctx.Input<Tensor>("W");
+
+    Tensor *out = ctx.Output<Tensor>("Out");
+    Tensor *last_h = ctx.Output<Tensor>("last_h");
+    Tensor *last_c = ctx.Output<Tensor>("last_c");
+
+    const T *x_data = x->data<T>();
+    const T *init_h_data = init_h->data<T>();
+    const T *init_c_data = init_c->data<T>();
+
+    const T *w_data = w->data<T>();
+
+    T *out_data = out->mutable_data<T>(ctx.GetPlace());
+    T *last_h_data = last_h->mutable_data<T>(ctx.GetPlace());
+    T *last_c_data = last_c->mutable_data<T>(ctx.GetPlace());
+
+    size_t max_len = ctx.Attr<int>("max_len");
+    float dropout_prob = ctx.Attr<float>("dropout_prob");
+    bool is_bidirec = ctx.Attr<bool>("is_bidirec");
+    int input_size = ctx.Attr<int>("input_size");
+    int hidden_size = ctx.Attr<int>("hidden_size");
+    int num_layers = ctx.Attr<int>("num_layers");
+    bool is_test = ctx.Attr<bool>("is_test");
+
+    /*
+    if (is_test) {
+      TensorCopy(*x, ctx.GetPlace(), out);
+      return;
+    }*/
+
+    auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
+    auto *cache_var = ctx.InputVar("Cache");
+    if (!cache_var) {
+      // The RAW type cache variable wouldn't be created and broadcasted on
+      // multi-devices before the first running.
+      // use parent scope to make cache persistable
+      auto *scope = const_cast<framework::Scope *>(ctx.scope().parent());
+      auto cache_var_name = ctx.Inputs("Cache")[0];
+      cache_var = scope->Var(cache_var_name);
+    }
+    CudnnRNNCache *cudnn_rnn_cache = nullptr;
+    if (cache_var->IsInitialized()) {
+      cudnn_rnn_cache = const_cast<framework::Variable *>(cache_var)
+                            ->GetMutable<CudnnRNNCache>();
+    } else {
+      cudnn_rnn_cache = const_cast<framework::Variable *>(cache_var)
+                            ->GetMutable<CudnnRNNCache>();
+      std::random_device rnd;
+      int seed = ctx.Attr<int>("seed");
+      if (seed == -1) {
+        seed = rnd();
+      }
+
+      auto input_w_numel = w->numel();
+      auto batch_size = x->dims()[1];
+      cudnn_rnn_cache->init(handle, ctx, max_len, batch_size, input_size,
+                            hidden_size, num_layers, dropout_prob, is_bidirec,
+                            seed, input_w_numel);
+    }
+
+    auto run_seq_len = x->dims()[0];
+
+    if (is_test) {
+      // for inference
+      CUDNN_ENFORCE(platform::dynload::cudnnRNNForwardInference(
+          handle, cudnn_rnn_cache->rnn_desc_, run_seq_len,
+          cudnn_rnn_cache->x_desc_, x_data, cudnn_rnn_cache->hx_desc_,
+          init_h_data, cudnn_rnn_cache->cx_desc_, init_c_data,
+          cudnn_rnn_cache->w_desc_, w_data, cudnn_rnn_cache->y_desc_, out_data,
+          cudnn_rnn_cache->hy_desc_, last_h_data, cudnn_rnn_cache->cy_desc_,
+          last_c_data, cudnn_rnn_cache->workspace_data_.data<uint8_t>(),
+          cudnn_rnn_cache->workspace_size_));
+    } else {
+      // for train
+      CUDNN_ENFORCE(platform::dynload::cudnnRNNForwardTraining(
+          handle, cudnn_rnn_cache->rnn_desc_, run_seq_len,
+          cudnn_rnn_cache->x_desc_, x_data, cudnn_rnn_cache->hx_desc_,
+          init_h_data, cudnn_rnn_cache->cx_desc_, init_c_data,
+          cudnn_rnn_cache->w_desc_, w_data, cudnn_rnn_cache->y_desc_, out_data,
+          cudnn_rnn_cache->hy_desc_, last_h_data, cudnn_rnn_cache->cy_desc_,
+          last_c_data, cudnn_rnn_cache->workspace_data_.data<uint8_t>(),
+          cudnn_rnn_cache->workspace_size_,
+          cudnn_rnn_cache->reserve_data_.data<uint8_t>(),
+          cudnn_rnn_cache->reserve_size_));
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class CudnnLSTMGPUGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *input = ctx.Input<Tensor>("Input");
+    auto *weight = ctx.Input<Tensor>("W");
+    auto *init_h = ctx.Input<Tensor>("InitH");
+    auto *init_c = ctx.Input<Tensor>("InitC");
+    // auto * last_h = ctx.Input<Tensor>("last_h");
+    // auto * last_c = ctx.Input<Tensor>("last_c");
+    auto *out = ctx.Input<Tensor>("Out");
+    auto *out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto *last_h_grad = ctx.Input<Tensor>(framework::GradVarName("last_h"));
+    auto *last_c_grad = ctx.Input<Tensor>(framework::GradVarName("last_c"));
+
+    // auto* init_h = ctx.Input<Tensor>("init_h");
+    // auto* init_c = ctx.Input<Tensor>("init_c");
+
+    auto *in_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
+    auto *weight_grad = ctx.Output<Tensor>(framework::GradVarName("W"));
+    auto *init_h_grad = ctx.Output<Tensor>(framework::GradVarName("InitH"));
+    auto *init_c_grad = ctx.Output<Tensor>(framework::GradVarName("InitC"));
+
+    auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
+    auto handle = dev_ctx.cudnn_handle();
+    auto *cache_var = ctx.InputVar("Cache");
+    PADDLE_ENFORCE(cache_var->IsInitialized());
+    CudnnRNNCache *cudnn_rnn_cache =
+        const_cast<framework::Variable *>(cache_var)
+            ->GetMutable<CudnnRNNCache>();
+
+    auto input_dims = input->dims();
+    auto weight_dims = weight->dims();
+    auto init_h_dims = init_h->dims();
+    auto init_c_dims = init_c->dims();
+    in_grad->mutable_data<T>(ctx.GetPlace());
+    weight_grad->mutable_data<T>(ctx.GetPlace());
+    math::SetConstant<DeviceContext, T> zero;
+    zero(dev_ctx, in_grad, static_cast<T>(0.0));
+    zero(dev_ctx, weight_grad, static_cast<T>(0.0));
+
+    T *init_h_grad_data = NULL;
+    if (init_h_grad == nullptr) {
+      Tensor init_h_grad_temp;
+      init_h_grad_temp.mutable_data<T>(init_h_dims, ctx.GetPlace());
+      zero(dev_ctx, &init_h_grad_temp, static_cast<T>(0.0));
+
+      init_h_grad_data = init_h_grad_temp.data<T>();
+    } else {
+      init_h_grad->mutable_data<T>(init_h_dims, ctx.GetPlace());
+      zero(dev_ctx, init_h_grad, static_cast<T>(0.0));
+      init_h_grad_data = init_h_grad->data<T>();
+    }
+
+    T *init_c_grad_data = NULL;
+    if (init_c_grad == nullptr) {
+      Tensor init_c_grad_temp;
+      init_c_grad_temp.mutable_data<T>(init_c_dims, ctx.GetPlace());
+      zero(dev_ctx, &init_c_grad_temp, static_cast<T>(0.0));
+
+      init_c_grad_data = init_c_grad_temp.data<T>();
+    } else {
+      init_c_grad->mutable_data<T>(init_c_dims, ctx.GetPlace());
+      zero(dev_ctx, init_c_grad, static_cast<T>(0.0));
+      init_c_grad_data = init_c_grad->data<T>();
+    }
+
+    const T *last_h_grad_data = NULL;
+    if (last_h_grad == nullptr) {
+      Tensor last_h_grad_temp;
+      last_h_grad_temp.mutable_data<T>(init_h_dims, ctx.GetPlace());
+      zero(dev_ctx, &last_h_grad_temp, static_cast<T>(0.0));
+
+      last_h_grad_data = (const T *)last_h_grad_temp.data<T>();
+    } else {
+      last_h_grad_data = last_h_grad->data<T>();
+    }
+
+    const T *last_c_grad_data = NULL;
+    if (last_c_grad == nullptr) {
+      Tensor last_c_grad_temp;
+      last_c_grad_temp.mutable_data<T>(init_c_dims, ctx.GetPlace());
+      zero(dev_ctx, &last_c_grad_temp, static_cast<T>(0.0));
+
+      last_c_grad_data = (const T *)last_c_grad_temp.data<T>();
+    } else {
+      last_c_grad_data = last_c_grad->data<T>();
+    }
+
+    const T *out_grad_data = NULL;
+    if (out_grad == nullptr) {
+      Tensor out_grad_temp;
+      out_grad_temp.mutable_data<T>(out->dims(), ctx.GetPlace());
+      zero(dev_ctx, &out_grad_temp, static_cast<T>(0.0));
+
+      out_grad_data = (const T *)out_grad_temp.data<T>();
+    } else {
+      out_grad_data = out_grad->data<T>();
+    }
+
+    // zero( dev_ctx, last_h_grad, static_cast<T>(0.0));
+    // zero( dev_ctx, last_c_grad, static_cast<T>(0.0));
+
+    auto out_data = out->data<T>();
+    // auto out_grad_data = out_grad->data<T>();
+    auto weight_data = weight->data<T>();
+    auto init_h_data = init_h->data<T>();
+    auto init_c_data = init_c->data<T>();
+    auto in_grad_data = in_grad->data<T>();
+
+    auto work_data = cudnn_rnn_cache->workspace_data_.data<uint8_t>();
+    auto reserve_data = cudnn_rnn_cache->reserve_data_.data<uint8_t>();
+
+    auto run_seq_len = input_dims[0];
+    PADDLE_ENFORCE_LE((size_t)run_seq_len, cudnn_rnn_cache->max_length_,
+                      "cudnn running seq_len CAN not greater max_lengh");
+    CUDNN_ENFORCE(platform::dynload::cudnnRNNBackwardData(
+        handle, cudnn_rnn_cache->rnn_desc_, run_seq_len,
+        cudnn_rnn_cache->y_desc_, out_data, cudnn_rnn_cache->dy_desc_,
+        out_grad_data, cudnn_rnn_cache->dhy_desc_, last_h_grad_data,
+        cudnn_rnn_cache->dcy_desc_, last_c_grad_data, cudnn_rnn_cache->w_desc_,
+        weight_data, cudnn_rnn_cache->hx_desc_, init_h_data,
+        cudnn_rnn_cache->cx_desc_, init_c_data, cudnn_rnn_cache->dx_desc_,
+        in_grad_data, cudnn_rnn_cache->dhx_desc_, init_h_grad_data,
+        cudnn_rnn_cache->dcx_desc_, init_c_grad_data, work_data,
+        cudnn_rnn_cache->workspace_size_, reserve_data,
+        cudnn_rnn_cache->reserve_size_));
+
+    CUDNN_ENFORCE(platform::dynload::cudnnRNNBackwardWeights(
+        handle, cudnn_rnn_cache->rnn_desc_, run_seq_len,
+        cudnn_rnn_cache->x_desc_, input->data<T>(), cudnn_rnn_cache->hx_desc_,
+        init_h->data<T>(), cudnn_rnn_cache->y_desc_, out->data<T>(),
+        cudnn_rnn_cache->workspace_data_.data<uint8_t>(),
+        cudnn_rnn_cache->workspace_size_, cudnn_rnn_cache->dw_desc_,
+        weight_grad->data<T>(), cudnn_rnn_cache->reserve_data_.data<uint8_t>(),
+        cudnn_rnn_cache->reserve_size_));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    cudnn_lstm,
+    ops::CudnnLSTMGPUKernel<paddle::platform::CUDADeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    cudnn_lstm_grad,
+    ops::CudnnLSTMGPUGradKernel<paddle::platform::CUDADeviceContext, float>);

paddle/fluid/operators/lstm_cudnn_op.h

+/* 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. */
+
+#pragma once
+#include <string>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/detail/activation_functions.h"
+#include "paddle/fluid/operators/math/lstm_compute.h"
+#include "paddle/fluid/operators/math/sequence2batch.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+class CudnnLSTMKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_THROW(
+        "CPU is not support for this kernel now. Will be add in the future");
+  }
+};
+
+template <typename DeviceContext, typename T>
+class CudnnLSTMGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {}
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/platform/dynload/cudnn.h

@@ -111,7 +111,23 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
   __macro(cudnnFindConvolutionForwardAlgorithmEx);        \
   __macro(cudnnFindConvolutionBackwardFilterAlgorithmEx); \
   __macro(cudnnFindConvolutionBackwardDataAlgorithmEx);   \
-  __macro(cudnnGetErrorString);
+  __macro(cudnnGetErrorString);                           \
+  __macro(cudnnCreateDropoutDescriptor);                  \
+  __macro(cudnnDropoutGetStatesSize);                     \
+  __macro(cudnnSetDropoutDescriptor);                     \
+  __macro(cudnnCreateRNNDescriptor);                      \
+  __macro(cudnnSetRNNDescriptor);                         \
+  __macro(cudnnGetRNNParamsSize);                         \
+  __macro(cudnnGetRNNWorkspaceSize);                      \
+  __macro(cudnnGetRNNTrainingReserveSize);                \
+  __macro(cudnnRNNForwardTraining);                       \
+  __macro(cudnnRNNBackwardData);                          \
+  __macro(cudnnRNNBackwardWeights);                       \
+  __macro(cudnnRNNForwardInference);                      \
+  __macro(cudnnDestroyDropoutDescriptor);                 \
+  __macro(cudnnDestroyRNNDescriptor);                     \
+  __macro(cudnnSetRNNDescriptor_v6);
+
 CUDNN_DNN_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 
 #define CUDNN_DNN_ROUTINE_EACH_R2(__macro) \

python/paddle/fluid/layers/nn.py

@@ -169,6 +169,7 @@ __all__ = [
     'log_loss',
     'add_position_encoding',
     'bilinear_tensor_product',
+    'lstm',
 ]
 
 
@@ -472,6 +473,169 @@ def dynamic_lstm(input,
     return hidden, cell
 
 
+def lstm(input,
+         init_h,
+         init_c,
+         max_len,
+         dropout_prob,
+         input_size,
+         hidden_size,
+         num_layers,
+         is_bidirec=False,
+         dtype='float32',
+         is_test=False,
+         name=None,
+         default_initializer=None,
+         seed=-1):
+    """
+    If Device is GPU, This op will use cudnn LSTM implementation
+
+    A four-gate Long Short-Term Memory network with no peephole connections.
+    In the forward pass the output ht and cell output ct for a given iteration can be computed from the recurrent input ht-1, 
+    the cell input ct-1 and the previous layer input xt given matrices W, R and biases bW, bR from the following equations:
+
+    $$ i_t = \\sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + bx_i + bh_i) $$
+
+    $$ f_t = \\sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + bx_f + bh_f) $$
+
+    $$ o_t = \\sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + bx_o + bh_o) $$
+
+    $$ \\tilde{c_t} = tanh(W_{cx}x_t + W_{ch}h_{t-1} + bx_c + bh_c) $$
+
+    $$ c_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} $$
+
+    $$ h_t = o_t \\odot tanh(c_t) $$
+
+    - W terms denote weight matrices (e.g. $W_{ix}$ is the matrix
+      of weights from the input gate to the input)
+    - The b terms denote bias vectors ($bx_i$ and $bh_i$ are the input gate bias vector).
+    - sigmoid is the logistic sigmoid function.
+    - $i, f, o$ and $c$ are the input gate, forget gate, output gate,
+      and cell activation vectors, respectively, all of which have the same size as
+      the cell output activation vector $h$.
+    - The $\odot$ is the element-wise product of the vectors.
+    - `tanh` is the activation functions.
+    - $\tilde{c_t}$ is also called candidate hidden state,
+      which is computed based on the current input and the previous hidden state.
+
+    Where sigmoid is the sigmoid operator: sigmoid(x) = 1 / (1 + e^-x), * represents a point-wise multiplication, 
+    X represensts a matrix multiplication
+
+
+    Args:
+        input (Variable): LSTM input tensor, shape MUST be ( seq_len x batch_size x input_size )
+        init_h(Variable): The initial hidden state of the LSTM                       
+                       This is a tensor with shape ( num_layers x batch_size x hidden_size)
+                       if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
+        init_c(Variable): The initial cell state of the LSTM.
+                       This is a tensor with shape ( num_layers x batch_size x hidden_size )
+                       if is_bidirec = True, shape should be ( num_layers*2 x batch_size x hidden_size)
+        max_len (int): max length of LSTM. the first dim of input tensor CAN NOT greater than max_len 
+        dropout_prob(float): dropout prob, dropout ONLY work between rnn layers, NOT between time steps
+                             There is NO dropout work on rnn output of the last RNN layers
+        input_size (int): hidden size of the input tensor
+        hidden_size (int): hidden size of the LSTM
+        num_layers (int): total layers number of the LSTM
+        is_bidirec (bool): If it is bidirectional
+        dtype (str): Data type. Choices = ["float32", "float64"], default "float32".
+        is_test (bool): If it is in test phrase
+        name (str|None): A name for this layer(optional). If set None, the layer
+                         will be named automatically.
+        default_initializer(Initialize|None): Where use initializer to initialize the Weight
+                         If set None, defaule initializer will be used
+        seed(int): Seed for dropout in LSTM, If it's -1, dropout will use random seed
+
+
+    Returns:
+        rnn_out(Tensor): result of LSTM hidden, shape is (seq_len x batch_size x hidden_size)
+                         if is_bidirec set to True, shape will be ( seq_len x batch_sze x hidden_size*2)
+        last_h(Tensor): the hidden state of the last step of LSTM
+                        shape is ( num_layers x batch_size x hidden_size )
+                        if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size)                     
+        last_c(Tensor): the cell state of the last step of LSTM
+                        shape is ( num_layers x batch_size x hidden_size )
+                        if is_bidirec set to True, shape will be ( num_layers*2 x batch_size x hidden_size)                     
+
+
+    Examples:
+        .. code-block:: python
+
+            input = embedding
+            batch_size = 20
+            max_len = 100
+            dropout_prob = 0.2
+            input_size = 100
+            hidden_size = 150
+            num_layers = 1
+            init_hidden1 = layers.fill_constant( [num_layers, batch_size, hidden_size], 'float32', 0.0, stop_grad=False)
+            init_cell1 = layers.fill_constant( [num_layers, batch_size, hidden_size], 'float32', 0.0, stop_grad=False)
+
+            rnn_out, last_h, last_c = layers.lstm( input, init_h, init_c, \
+                    max_len, dropout_prob, input_size, hidden_size, \
+                    num_layers)
+    """
+
+    helper = LayerHelper('cudnn_lstm', **locals())
+
+    weight_size = 0
+    for i in range(num_layers):
+        if i == 0:
+            input_weight_size = (input_size * hidden_size) * 4
+        else:
+            if is_bidirec:
+                input_weight_size = (hidden_size * 2 * hidden_size) * 4
+            else:
+                input_weight_size = (hidden_size * hidden_size) * 4
+
+        hidden_weight_size = (hidden_size * hidden_size) * 4
+
+        if is_bidirec:
+            weight_size += (input_weight_size + hidden_weight_size) * 2
+            weight_size += hidden_size * 8 * 2
+        else:
+            weight_size += input_weight_size + hidden_weight_size
+            weight_size += hidden_size * 8
+
+    weight = helper.create_parameter(
+        attr=helper.param_attr,
+        shape=[weight_size],
+        dtype=dtype,
+        default_initializer=default_initializer)
+
+    out = helper.create_variable_for_type_inference(dtype)
+    last_h = helper.create_variable_for_type_inference(dtype)
+    last_c = helper.create_variable_for_type_inference(dtype)
+
+    cache = helper.create_variable(
+        persistable=True, type=core.VarDesc.VarType.RAW, stop_gradient=True)
+
+    helper.append_op(
+        type='cudnn_lstm',
+        inputs={
+            'Input': input,
+            'InitH': init_h,
+            'InitC': init_c,
+            'W': weight,
+            'Cache': cache,
+        },
+        outputs={
+            'Out': out,
+            'last_h': last_h,
+            'last_c': last_c,
+        },
+        attrs={
+            'max_len': max_len,
+            'is_bidirec': is_bidirec,
+            'input_size': input_size,
+            'hidden_size': hidden_size,
+            'num_layers': num_layers,
+            'is_test': is_test,
+            'dropout_prob': dropout_prob,
+            'seed': seed,
+        })
+    return out, last_h, last_c
+
+
 def dynamic_lstmp(input,
                   size,
                   proj_size,